TW503375B - Method for hyperspectral imagery exploitation and pixel spectral unmixing - Google Patents

Method for hyperspectral imagery exploitation and pixel spectral unmixing Download PDF

Info

Publication number
TW503375B
TW503375B TW089108706A TW89108706A TW503375B TW 503375 B TW503375 B TW 503375B TW 089108706 A TW089108706 A TW 089108706A TW 89108706 A TW89108706 A TW 89108706A TW 503375 B TW503375 B TW 503375B
Authority
TW
Taiwan
Prior art keywords
module
pixel
content
fuzzy
binary
Prior art date
Application number
TW089108706A
Other languages
Chinese (zh)
Inventor
Ching-Fang Lin
Original Assignee
Ching-Fang Lin
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Priority claimed from US09/351,349 external-priority patent/US6665438B1/en
Application filed by Ching-Fang Lin filed Critical Ching-Fang Lin
Application granted granted Critical
Publication of TW503375B publication Critical patent/TW503375B/en

Links

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/21Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G06F18/2134Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on separation criteria, e.g. independent component analysis

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Data Mining & Analysis (AREA)
  • Theoretical Computer Science (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Image Processing (AREA)

Abstract

An efficiently hybrid approach to exploit hyperspectral imagery and unmix spectral pixels. This hybrid approach uses a genetic algorithm to solve the abundance vector for the first pixel of a hyperspectral image cube. This abundance vector is used as initial state in a robust filter to derive the abundance estimate for the next pixel. By using Kalman filter, the abundance estimate for a pixel can be obtained in one iteration procedure which is much fast than genetic algorithm. The output of the robust filter is fed to genetic algorithm again to derive accurate abundance estimate for the current pixel. The using of robust filter solution as starting point of the genetic algorithm speeds up the evolution of the genetic algorithm. After obtaining the accurate abundance estimate, the procedure goes to next pixel, and uses the output of genetic algorithm as the previous state estimate to derive abundance estimate for this pixel using robust filter. And again use the genetic algorithm to derive accurate abundance estimate efficiently based on the robust filter solution. This iteration continues until pixels in a hyperspectral image cube end.

Description

〜3 ί) 〜3 ί) Α7 Β7 、發明説明( 高譜數據這備語有其難性。所有_難的共同 特徵是含有大量駭及窄光譜頻段。每健譜圖象的確切 頻段敦差別很大。位於可見光波長範圍内的單一波段可在 1納米到幾賴米範51峻化。位於紅外_波長範圍內 5的波段範圍可能超過可見光波長範圍內的波段。當然,高 譜數據是極其有用的,因為利用高譜數據及觀測到的物體 的極特別特徵很容易識別該物體。物體的這些極特別特徵 涉及到非4乍的頻譜波段。這種檢測與識別不易用傳統方 法實現。與高譜數據相關的缺點是要求處理龐大信息量的 10能力。特定的元素,物體或目榡,或成份,都擁有其特定 的,譜特徵。鑒^-個特定的頻譜特徵即可鑒別相對應的 兀素。現有處理高譜數據的方法包括模式匹配技術。這些 技術都依賴模型及最小二乘算法來識別與隔離高譜数據内 的元素。這些模式匹配技術囡其缺乏魯棒性而受,制。 15,結果個在時間與空間上的魏臓著下降。它們也不 含匕根據一個組合頻譜特徵來識別其元素的成份,且計算吾 龐大。其結果也因傳感器及大氣變化而顯著下降。傳統方 法也不能處理非線性1增加在高譜數據中檢測的元素數 據庫時,這些方法效果也不佳。 ^ -0 聽計算技術(Ε〇非常適合於非線性優化。進化計 算技術是近期發展的高級計算領域中進一步發展的結果: 高級計算領域包㈣傳算法,遺傳規劃,人工神經網^, 及人工生命。這些技術很適合移植到高譜數據及^式匹配 。除了匹配數據庫_頻譜特徵與從高_象提取的頻譜 (請先閱讀背面之注意事項再填寫本頁) 訂------線#-----~ 3 ί) ~ 3 ί) Α7 Β7, invention description (high spectrum data, the idiom has its difficulty. The common feature of all _ difficult is that it contains a lot of scary narrow spectrum frequency bands. The exact frequency band of each healthy spectrum image is different. Very large. A single band in the visible wavelength range can be sharpened from 1 nanometer to a few millimeters. 51 The band in the infrared_wavelength range 5 may exceed the band in the visible wavelength range. Of course, the hyperspectral data is extremely Useful because it is easy to identify the object using hyperspectral data and the extremely special characteristics of the observed object. These extremely special characteristics of the object relate to non-four-frequency spectrum bands. This detection and identification is not easy to achieve with traditional methods. The shortcomings related to high-spectrum data are the ability to process huge amounts of information. Specific elements, objects or objects, or components have their specific, spectral characteristics. You can identify the corresponding spectral characteristics by identifying a specific spectral characteristic. The existing methods for processing hyperspectral data include pattern matching techniques. These techniques rely on models and least squares algorithms to identify and isolate elements within hyperspectral data. These pattern matching techniques suffer from their lack of robustness. 15. As a result, Wei Wei's decline in time and space. They also do not include the use of a combination of spectral characteristics to identify the composition of its elements, and calculate the Huge. The results are also significantly reduced due to changes in sensors and the atmosphere. Traditional methods also cannot handle non-linear 1 and increase the database of elements detected in hyperspectral data. These methods are also not effective. ^ -0 Listening computing technology (Ε〇 It is very suitable for non-linear optimization. Evolutionary computing technology is the result of further development in the recently developed advanced computing field: advanced computing fields include algorithms, genetic programming, artificial neural networks, and artificial life. These technologies are very suitable for transplantation to high-level Spectral data and ^ pattern matching. In addition to matching the database_spectrum characteristics and the spectrum extracted from the height_image (please read the precautions on the back before filling this page) Order ------ 线 # -----

83· 3· 10,000 、發明説明(2) 10 15 20 秦 特徵之外,這些進化計算技術也從遞推搜尋過程中產生一 些特別的方法,這些方法可用於分析高譜數據。這些技術 具有“學習”能力。進化算法的原始依據是虚擬物體環境 ,每個物體都表示用於分析一組數據特定方面的方法。進 化算法的描述參考第一圖。 達爾文關於遺傳與進化過程中的“適者生存,,的思想 是進化算法方法用於優化任務的基石。從解決一特定問題 的所有可能方法的集合出發,新一代獨特且預期更好(或 更適合)的方法將從其父代或原方法的隨意配對中產生。 一旦父代方法被選出並配對後,它們將相互交換部分各自 的方法。方法或染色體的交換稱為交配或繁殖。兩個單獨 的原方法從繁殖過程中產生,且父代方法將不再存在。只 有其下一代從交配過程生存下來。下一代方法一定能分析 其父代能處理的同一方面的數據。變異可能隨機出現在每 個後續代中以引入必要的多樣性。 每個後續代的子方法將進行其處理能力的測試。用户 設置一定指標對子方法進行評定。每一代的方法都得到一 個分數。依據用户的指標,這些分數用於指出在分析數據 中個方法的適應度,或避合度<5低適合度的方法不許配 對及繁衍後代,這樣它們逐渐滅亡,不再存活。那垄具有 高適合度分數的方法將允許配對及交配,囡而,它們將繁 衍自己並延續其種族。這種適者生存進化方法延續特定的 ,用户選定的代數。這導致分析特定數據集合的優化方法 快速收斂。諸如頻譜分解,目標檢測與識別這類任務極適 本紙張从逍用中賴家揉準(CNS ) 44胁(210X297公釐 83.3.10,000 (請先閣讀背面之注意事項再填寫本頁) 503375 A7 、發明説明(3 10 15 1 20 法 合這類進化方法。因此,作為以上論逑的延續,高譜數據 是進化計算技術的一個理想應用對象。 在經歷所有允許的代數後,在過程結束時,根據用户 指標評判得到的具有最好適合度分數的方法也就是分析給 定數據集合的最優方法。一般説來,進化算法給出比傳統 方法好得多的結果,因為搜索過程具有很多的選擇性。進 化算法同時在整個候選方法集合中搜尋。進化算法沒有任 何困難進行多約束的複雜優化任務。對非線性也沒有過多 的限制什麼問題。非線性常常體現在與高譜圖象數據相關 的大氣與傳感器约束條件上。與適應度評定相關聯的指標 函數也反映了這些非線性與約束條件。 第二圖反映了進化計算技術在高譜數據中的應用。該 應用包括一項處理與頻譜波長相關的單個像素特徵的可靠 技術。進化算法技術包括特徵頂處理以及基於高譜圖象數 據的模型約朿的非參數搜尋。 當應用進化計算技術時,高譜圖象的每個像素需要進 行成千上萬次的遞推運算。這個要求導致非常慢的計算遇 程。一個適當大小的高譜圖象需要幾百萬次的遞推運算。 因此,需要一個對高譜圖象分析和精確分解的快速處理方 緣是,為達到上述目的,本發明提供一種混合方法, 該方法利用魯棒濾波技術來進行快速高譜圖象的像素分解 ,並利用遺傳算法來細化魯棒濾波器的含量估計。 本發明之另一目標是提供一種混合方法,當魯棒濾波 -5㈣ 本紙張ΛΑϋ用 t ® S 家縣(CNS) A4*L^( 210x 297/^¾ g3. 3. !0,〇〇〇 (請先閲讀背面之注意事項再填寫本頁)83 · 3 · 10,000, Invention Description (2) 10 15 20 Qin In addition to these characteristics, these evolutionary computing techniques also generate some special methods from the recursive search process, which can be used to analyze hyperspectral data. These technologies are “learning”. The original basis of evolutionary algorithms is the environment of virtual objects, each of which represents a method for analyzing a particular aspect of a set of data. For the description of the evolution algorithm, refer to the first figure. Darwin's idea of "survival of the fittest" in the process of genetics and evolution is the cornerstone of the evolutionary algorithm approach used to optimize tasks. Starting from the set of all possible methods to solve a particular problem, the new generation is unique and expected to be better (or more suitable) ) Method will be generated from the random pairing of its parent or original method. Once the parent method is selected and paired, they will exchange parts of their respective methods. The method or chromosome exchange is called mating or reproduction. Two separate The original method is generated from the breeding process, and the parent method will no longer exist. Only its next generation survives the mating process. The next generation method must be able to analyze the same data that its parent can handle. The mutation may appear randomly in The necessary diversity is introduced in each subsequent generation. The sub-method of each subsequent generation will be tested for its processing ability. The user sets a certain index to evaluate the sub-method. Each generation of the method gets a score. According to the user's index These scores are used to indicate the fitness of the methods in the analysis data, or the degree of avoidance < 5 low fitness The degree method does not allow pairing and breeding of offspring, so that they gradually perish and no longer survive. The method with a high fitness score will allow pairing and mating, but they will breed themselves and continue their race. The fittest survives and evolves The method continues to a specific, user-selected algebra. This leads to rapid convergence of optimization methods for analyzing specific data sets. Tasks such as spectral decomposition, target detection and recognition are extremely suitable for this paper. (210X297 mm 83.3.10,000 (please read the precautions on the back before filling out this page) 503375 A7, invention description (3 10 15 1 20 method combined with this type of evolutionary method. Therefore, as a continuation of the above argument, high-spectrum Data is an ideal application object of evolutionary computing technology. After experiencing all allowed algebras, at the end of the process, the method with the best fitness score obtained according to user indicators is the best method for analyzing a given data set. In general, evolutionary algorithms give much better results than traditional methods, because the search process is very selective. The search algorithm searches through the entire set of candidate methods at the same time. The evolutionary algorithm does not have any difficulty to perform complex optimization tasks with multiple constraints. There is no excessive restriction on non-linearity. Non-linearity is often reflected in the atmosphere and data related to hyperspectral image data. Sensor constraints. The index function associated with fitness assessment also reflects these non-linearities and constraints. The second figure reflects the application of evolutionary computing techniques in hyperspectral data. This application includes a process that is related to spectral wavelengths Reliable technology for single pixel features. Evolutionary algorithm technology includes feature top processing and non-parametric search of models based on hyperspectral image data. When applying evolutionary computation techniques, each pixel of a hyperspectral image needs to perform thousands of Tens of thousands of recursive operations. This requirement results in very slow computational strokes. An appropriately sized hyperspectral image requires millions of recursive operations. Therefore, what is needed is a fast processing method for the analysis and accurate decomposition of hyperspectral images. In order to achieve the above purpose, the present invention provides a hybrid method that uses robust filtering technology to perform fast pixel decomposition of hyperspectral images. And the genetic algorithm is used to refine the content estimation of the robust filter. Another object of the present invention is to provide a hybrid method, when robust filtering -5㈣ of this paper ΛΑϋ uses t ® S Home County (CNS) A4 * L ^ (210x 297 / ^ ¾ g3. 3.! 0, 〇〇〇〇 (Please read the notes on the back before filling this page)

、發明説明(4 A7 B7、 Explanation of invention (4 A7 B7

I 時,該方法利用魯棒濾波技術_ 丁 μ Ρ曰圖象的像素分解,並利用遺傳算法來得到 準確含量估計。 本發日f之另一目榡是提供-種混合方法,該方法利用 5魯,卡爾曼濾波器來進行快速高譜圖象的像素分解,並利 用遺傳算法來細化魯棒濾波器的含量估計。 本發明之另一目標是提供一種混合方法,當魯棒濾波 器的估P十誤差大於預設時,該方法利用魯棒卡爾曼濾波器 來進行快速高譜圖象的像素分解,並利用遺傳算法來得到 10 準確含量估計。 圖示説明 第一圖:是一個方塊圖,描述進化計算的方法。 第二圖:是一個方塊圖,描述進化計算方法用於高譜圖象 的頻譜分解方法。 15第三圖:是一個功能方塊圖,描述該混合方法的優選實現 。依據本發明,該混合方法用於高譜圖象分析與 像素分解。 第四圖:是一個功能方塊圖,描述該混合方法的第二優選 實現。依據本發明,該混合方法用於高譜圖象分 20 析與像素分解。 第五圖:是一個功能方塊圖,描述利用魯棒卡爾曼濾波器 來進行超級頻譜圖象像素分解的實現。 第六圖:是一個功能方塊圖,描述模糊邏辑模塊的實現。 第七圖:是一個功能方塊圖,描述於高譜圖象像素分解的 -6- 本纸張尺度適用中國國家揉準(CNS ) Α4規格(21〇X:297公釐) 83. 3.10,000 (請先閲讀背面之注意事項再填寫本頁) 線_ 503375 A7 B7 五、發明説明(5 ) 10 15 遺傳算法實現。 圖號説明= 10-高譜圖象傳感器 30-大氣補償模塊 50-頻譜庫 .61 -模糊邏辑模塊 612-模糊推理機構 614-模糊還原模塊 63- 協方差傳播模塊 64- 最優增益計算模塊 66-狀態矢量更新模塊 68-狀態矢量預測模塊 Ή-適合度計算模塊 73-繁殖模塊 75-變異模塊 77-二進制字符串 20-圖象登記模塊 40-特徵預處理模塊 60-魯棒濾波器 61Η莫糊化模塊 613-模糊規則庫 62-預處理模塊 64- 最優增益計算模塊 65- 協方差更新模塊 67-測量殘差計算模塊 70-遺傳算法分解模塊 72-解碼模塊 74-交配模塊 76-編碼模塊 80-人工神經網絡模塊 (請先閲讀背面之注意事項再填寫本頁) 標' f # 鈣 % % 裝 90-第一像素點含量估計模塊 本發明提供一種有效的混合方法來分析高譜圖象及分 解頻譜像素。該混合方法利用遺傳算法求解高譜圖象立方 體的第一像素點的含量矢量。該含量矢量用作卡爾曼濾波 器的初始值,利用卡爾曼濾波技術得出下一個像素點的含 量估計。利用卡爾曼濾波器,像素點的含量估計可在卡爾 曼濾波過程中得到,且卡爾曼濾波比遺傳算法快得多。 卡爾、曼濾波器;的输出再提供給遺傳算法來得出當前像 -7- 本紙ίΜ適财關家縣(CNS ) A4· ( 21Qx297公董) 83. 3. 10,000 503375 A7 B7 五、發明説明(6 ) 素點的準確含量估計。卡爾曼濾波器的解作為遺傳算法的 起始點將加快遺傳算法的進化。在得到準確含量估計後, 栘到下一像素點,並利用遺傳算法的輸出作為前一個狀態 估汁,利用卡爾曼濾波器得出當前點的含量估計。再基於 5卡爾曼解利用遗傳算法得出準確含量估計。依此類推,該 遞推過程一直進行到高譜圖象立方體中的最後一個像素點 〇 卡爾曼濾波器的思想是相對有效的。第一,假定一個 可由其狀態矢量描述的系統。在應用高譜圖象分析中 10 ,狀態矢量即是目標成員含量矢量。狀態矢量不能直接觀 測,但可以得到受噪音污染的觀測數據。該觀測數據 ’即像素點的測量值,通過下列測量方程與系統狀態相關 聯: zU)=h(xU))+v(々) 15 用索引值々來替代以表示在錄置的像素點。在 以上方程中,表示方差為此幻的測量噪音表示 矢量函數。該函數描述像素點測量值與含量矢量間的關係 假定有c種材料和^^!!(頻譜波段或通道。通常假定^大 20於或等於0以滿足可辨識性。根據高譜圖象的線性模型, 由一個矩陣表示。這樣就得到一個線性測量方程 逶 Ζ(女)=Sx〇)+v(々) 卡爾曼濾波是一個最小均方估計器,且有兩個明顯特 83. 3.10,000 (請先閱讀背面之注意事項再填寫本頁) 本紙張尺度適用中國國家揉準(CNS ) A4規格(210x297公瘦) 發明説明(7 10 徵。第—_徵是卡爾«、波是基於狀触纖念。該特 徵允許卡㈣毅將紐作為-健體來處理,而不是當 作-組單韻元件。錄是卡锻舰為遞推形式 。=態估計的紐a機當齡計和當祕緣據來計算 ^迫個特性使卡爾曼濾波比在濾波過程中的每一步都要整 個過去的输入數據來計算估計值要有效的多。 用卡爾曼濾波來分解高譜圖象的優點包括:(1)利用 相鄰像素點間的含量矢量關係來準確估計包含在一個像素 點里的材料百分比;(2)僅用一次卡爾曼遞推處理一個 像素點,這樣導致高效地分析高譜圖象。 在許多信息提取應用中,根據物理定律,不能由傳感 器直接測量的內部動態系統狀態的傳遞關係通常用一個非 線性運續時間微分方程描述·· (請先閲讀背面之注意事項再填寫本頁) IP" 訂 4Μ 漆 製 15 x( 〇=f(x( 0)+w( t) 其中,x(i)為狀態矢量,f(x)為描述系統動力學的非 線性矢量函數,w(0是協方差矩陣爲Q(〖)的系統噪音。 為對高譜圖象分解進行卡爾曼濾波,需要一個狀態方 程。這個方程必須把當前像素點的含量矢量與上一個 2〇像素點的含量矢量關聯起來。根據前一個方程,這 個關係描述如下: x(k+lh0(k+L k)x(k)\w(k) 其中,是CLC狀態轉移矩陣’並將在像素點 女的系統狀態與在像素點女"的系統狀態關聯起來。 本纸張又度速用中國國家橾準(CNS ) A4規格(加父的7公釐) 83.3.10,000 503375 A7 B7 五、發明説明(8 ) 10 15 $St,. 20 jsL· W 灘gn 參考第三圖,本發明的高譜圖象分解的混合方法由以 下步驟構成: (a) 從高譜圖象傳感器1〇接收一個高譜圖象立方體 。該高譜圖象立方體以波長及空間位置(不凡)表示一個場 景。 (b) 在圖象登記模塊2〇中,逐個波段登記圖象數據 。登記操作是“使一個波段的圖象轉換到另一個波段圖象 ”而不必涉及地圖坐標系變換。 這個步驟保證對應在一個波段圖象中的一個像素點的 物理位置與在另一個波段圖象中的像素點的物理位置一致 〇 (C)發送登記後的高譜圖象立方體到大氣補償模塊 30,在大氣補償模塊30內補償大氣效應。 在這一步中,大氣校正可用上市款體FLAASH (頻譜超 級立方體的快速視線大氣分析)來完成。FLAASH是一個基 於MODTRAN的“大氣校正”軟體包,由Hansc〇m空軍基地的 AirForcePhii1ipsLaboratoiy與SpectraiSciences,Inc· 開發。該軟體支持當前和規劃的紅外線一可見光――紫外線 高譜及多頻譜傳感器。該軟體提供了地表及大氣特性(諸 如地π反照率,地表咼度,水蒸汽,浮質與雲層的光學 深度,地表與温度)的準確物理學推導。 、(d)接收來自頻譜庫50的感興趣材料的頻譜特徵, 並在特徵預處理模塊4Q預處理這鉴特徵。 頻譜庫是一個已知材料與物體的數據庫,並提供其特 83· 3.10,000 (請先閲讀背面之注意事項再填寫本頁) 訂 -10- 503375 A7 B7 五、發明説明(9 徵 10 15 ,癢蹄#4¾¾¾¾ 20 运些從實驗室得到的特徵是對應單個材料相對波長的 反射率。 來自頻譜庫的特徵在特徵預處理模塊40中進行正交歸 一化處理。該過程將特徵空間分解成一組正交矢量集合。 它們是已知材料或物體的線性組合。 5(e)第一像素點含量估計模塊9〇接收來自特徵預處 理模塊40的正交特徵以及來自大氣補償模塊3〇的第一像素 點測量數據,並計算第一像素點的含量矢量。 第一像素點的含量估計將被用作魯棒濾波器6〇的初始 值。除第一像素點以外的像素點,第一像素點含量估計模 塊90則繞過來自大氣補償模塊3〇的測量數據。 有不同的算法可用於估計第一像素點的含量矢量。可 選用的算法包括最小二乘(LS)估計器,極大似然法(ML )’及進化算法(EA)。 (f) 魯棒濾波器接收來自特徵預處理模塊4〇的正交 特徵,來自大氣補償模塊30的當前像素點測量數據,以及 來自第一像素點的含量矢量估計模塊9〇的含量估計,並利 用卡爾曼濾波技術進行當前像素點的頻譜分解。魯棒濾波 器60输出當前像素點的含量估計給遺傳算法分解模塊7〇。 (g) 遺傳算法分解模塊70接收特徵預處理模塊40的 正交特徵,來自大氣補償模塊30的當前像素點測量值,以 及來自魯棒濾波器60的當前像素點的含量估計,並利用遺 傳算法進行當前像素點的準確頻譜分解以得出準確含量估 計。 -11- ---------- (請先閲讀背面之注意事項再填寫本頁) -訂 本紙張尺度適用中國國家揉準(CNS ) A4規格(210X297公釐) 83·3·丨〇,0〇〇 503375 A7 B7 五、發明説明(10) 該準確含量估計即為.系統輸出。同時,該值也反饋到 魯棒濾波器60在魯棒濾波器中作為對下一個像素點的含量 估計的新初始值。 在這一步中,魯棒濾波器60的输出用作遺傳算法分解 5模塊7〇的起始點,以加速遺傳算法的進北通程/·- 、(h)移到下一個像素點,魯棒濾波器60接收來自特 •徵預處理模塊40的正交特徵,來自大氣補償模塊3〇的當前 像素點測量值,以及來自遺傳分解模塊70的前一個像素點 的含量估計,並利用卡爾曼濾波技術對當前像素點進行頻 10譜分解。在這一步中,來自遺傳算法分解模塊7〇的準確含 量估計在魯棒濾波器60中用作為前一個像素點的含量估計 以便準確估計當前像素點的含量矢量。 (i)回到步驟(g),在高譜特性立方體内重復遞推步 驟(g)與(h)直到最後一個像素點。 15 值侍指出的是魯棒濾波器60與遺傳算法可以以並行方 式運行。 如第四圖所示,在步驟(d)之後,加入了另一個步驟 ,即利用人工神經網絡模塊80進行材料分類選擇。這一步 可减少含量矢量的元素以達到加速高譜特性分析^因此, 20在步驟(d)之後,本發明的方法由以下步驟組成。 (e’)人工神經網絡8〇接收來自特徵預處理模塊4〇的 正交特徵以及來自大氣補償模塊30的當前像素點測量數據 ,並進行材料分類選擇。分類數據送給遺傳算法分解模塊 70和魯棒濾波器6〇。第一像素點的分類數據也送給第一像 -12- 本纸财關家縣(CNS) A4· ( 21GX297公董) 83.3.10,000 (請先閲讀背面之注意事項再填寫本頁) 訂 五 、發明説明(ll A7 B7 J點含量佩购0。_,人工神經_她繞過從大 氣補償模塊30到後續模塊的測量數據。 、來自特顏處觀塊4G的JE交絲(酿為—)稱之 =類,並被人I神經網絡_於校正傳感器_。來自 當前補償模塊30的校正後的像素點輸入數據由人工神經網 絡處理以健錄讀徵與麵賴已知分難徵之^的 相關性程度。 ★在模糊神經網絡中,所有神經元的输出用於評估以確 定是否有神經元的響應超過預設值,比如·〇· 5 (神經元的 输出範圍:0·0〜1·0) β如果在神經網絡中有一個或多個 神經兀的響應超過預設值,對應該神經元的分類被選作用 =續處_候選方法。分類選簡優點是只有與當前像 素點相關的材料含量才在卡爾曼濾波器和遺傳算法中用作 為狀態矢量的元素。這一步可減少含量矢量的元素個數, 15 達到加速處理。 (f’)第一像素點含量估計模塊90接收來自特徵預處 理模塊的JE交特徵,來自大㈣償模塊3_第—像素點的 測量數據,以及來自人工神經網絡8〇的對第一像素點的分 類數據,並計算第一像素點的含量矢量。第一像素點的含 量估計值用作為魯棒濾波器6〇的初始值。對於除去第一像 素點以外的像素點,第一像素點含量估計模塊90僅僅將繞 過來自人工神經網絡80的測量數據。 有不同的算法可用於第一像素點的含量矢量估計 選用的算法包括最小二乘(LS)佑計器,極大似然法 10 20 可At I, this method uses robust filtering techniques. Ding Ping refers to the pixel decomposition of the image, and uses genetic algorithms to obtain accurate content estimates. Another goal of this issue is to provide a hybrid method, which uses 5 Lu, Kalman filters to perform fast pixel decomposition of hyperspectral images, and uses genetic algorithms to refine the content estimation of robust filters. . Another object of the present invention is to provide a hybrid method, which uses a robust Kalman filter to perform pixel decomposition of fast hyperspectral images when the estimated P 十 error of the robust filter is greater than a preset, and uses genetic Algorithm to get 10 accurate content estimates. Graphical illustrations First figure: is a block diagram describing the method of evolutionary computation. Figure 2: A block diagram depicting the spectral decomposition method of evolutionary computation for hyperspectral images. 15 The third figure: is a functional block diagram describing the preferred implementation of the hybrid method. According to the invention, the hybrid method is used for hyperspectral image analysis and pixel decomposition. Figure 4: A functional block diagram depicting a second preferred implementation of the hybrid approach. According to the present invention, the hybrid method is used for hyperspectral image analysis and pixel decomposition. Figure 5: A functional block diagram depicting the implementation of pixel decomposition of a super-spectral image using a robust Kalman filter. Figure 6: A functional block diagram describing the implementation of the fuzzy logic module. The seventh figure: is a functional block diagram, described in the pixel decomposition of the hyperspectral image. -6- This paper size is applicable to the Chinese National Standard (CNS) A4 specification (21〇X: 297 mm) 83. 3.10,000 (Please read the precautions on the back before filling out this page) Line_ 503375 A7 B7 V. Description of the invention (5) 10 15 Realization of genetic algorithm. Explanation of drawing number = 10- hyperspectral image sensor 30- atmospheric compensation module 50- spectrum library. 61-fuzzy logic module 612-fuzzy inference mechanism 614-fuzzy reduction module 63-covariance propagation module 64-optimal gain calculation module 66- state vector update module 68- state vector prediction module Ή- fitness calculation module 73- reproduction module 75- mutation module 77- binary string 20- image registration module 40- feature pre-processing module 60- robust filter 61 Η Modification module 613-Fuzzy rule base 62-Preprocessing module 64-Optimal gain calculation module 65-Covariance update module 67-Measurement residual calculation module 70-Genetic algorithm decomposition module 72-Decoding module 74-Mating module 76- Coding module 80-artificial neural network module (please read the notes on the back before filling this page) marked 'f # calcium%% installed 90- first pixel content estimation module The present invention provides an effective hybrid method to analyze the high spectrum Image and resolution spectrum pixels. This hybrid method uses a genetic algorithm to solve the content vector of the first pixel of the hyperspectral image cube. The content vector is used as the initial value of the Kalman filter, and the content estimation of the next pixel is obtained by using the Kalman filter technology. With the Kalman filter, the pixel content estimation can be obtained in the Kalman filtering process, and the Kalman filtering is much faster than the genetic algorithm. Kalman and Mann filter; the output is then provided to the genetic algorithm to obtain the current image -7- this paper ίΜ 财 关 家 县 (CNS) A4 · (21Qx297 public director) 83. 3. 10,000 503375 A7 B7 V. Description of the invention 6) Accurate content estimation of prime points. The Kalman filter solution as the starting point of the genetic algorithm will accelerate the evolution of the genetic algorithm. After the accurate content estimation is obtained, the next pixel point is taken, and the output of the genetic algorithm is used as the previous state estimation juice, and the content estimation of the current point is obtained by using the Kalman filter. Based on the 5 Kalman solution, the genetic algorithm is used to obtain an accurate content estimate. By analogy, the recursive process proceeds to the last pixel in the hyperspectral image cube. The idea of the Kalman filter is relatively effective. First, suppose a system that can be described by its state vector. In applying hyperspectral image analysis, 10, the state vector is the target member content vector. State vectors cannot be directly observed, but observation data contaminated by noise can be obtained. The observation data ′ is the measured value of the pixel, which is related to the system state by the following measurement equation: zU) = h (xU)) + v (々) 15 The index value 替代 is used to represent the pixel points recorded. In the above equation, the measurement noise representing the variance is a vector function representing the magic noise. This function describes the relationship between pixel measurement and content vector. It is assumed that there are c materials and ^^ !! (spectral band or channel. It is generally assumed that ^ is greater than 20 or equal to 0 to meet discernibility. According to the hyperspectral image's The linear model is represented by a matrix. In this way, a linear measurement equation 逶 Z (female) = Sx〇) + v (々) is obtained. The Kalman filter is a minimum mean square estimator, and has two obvious features: 83. 3.10, 000 (Please read the notes on the back before filling out this page) This paper size is applicable to the Chinese National Standard (CNS) A4 size (210x297 male thin) Invention description (7 10 sign. The first _ sign is Carl «, the wave is based on This feature allows Qiaoyi to treat the button as a-body, rather than as a single-single rhyme component. The record is a recursive form of the card forging ship. = State estimation of the button age machine Calculating the characteristics based on secret factors makes the Kalman filter more effective than the entire input data at each step in the filtering process to calculate the estimated value. Kalman filtering is used to decompose the hyperspectral image. The advantages include: (1) using the content vector between adjacent pixels Relationship to accurately estimate the percentage of material contained in a pixel; (2) processing a pixel with Kalman recursion only once, which leads to efficient analysis of hyperspectral images. In many information extraction applications, according to the laws of physics, The transfer relationship of the state of the internal dynamic system that cannot be directly measured by the sensor is usually described by a non-linear lifetime differential equation. (Please read the precautions on the back before filling this page) IP " Order 4M lacquer 15 x (〇 = f (x (0) + w (t) where x (i) is a state vector, f (x) is a non-linear vector function describing system dynamics, and w (0 is a system with a covariance matrix of Q (〖) Noise. In order to perform Kalman filtering on hyperspectral image decomposition, an equation of state is required. This equation must correlate the content vector of the current pixel with the content vector of the previous 20 pixel. According to the previous equation, this relationship describes As follows: x (k + lh0 (k + L k) x (k) \ w (k) where is the CLC state transition matrix 'and associates the system state at the pixel with the system state at the pixel. Get up. China National Standards (CNS) A4 specification (plus 7 mm of the parent) 83.3.10,000 503375 A7 B7 V. Description of the invention (8) 10 15 $ St ,. 20 jsL · W Tan gn The hybrid method of hyperspectral image decomposition is composed of the following steps: (a) A hyperspectral image cube is received from a hyperspectral image sensor 10. The hyperspectral image cube represents a scene in wavelength and spatial position (unusual). (B) In the image registration module 20, image data is registered on a band-by-band basis. The registration operation is to "convert an image in one band to another band image" without having to involve map coordinate system transformation. This step ensures that the physical position of a pixel corresponding to a band image is consistent with the physical location of a pixel in another band image. (C) Send the registered hyperspectral image cube to the atmospheric compensation module 30 The atmospheric effect is compensated in the atmospheric compensation module 30. In this step, atmospheric correction can be done with the market model FLAASH (Fast Sight Atmospheric Analysis of Spectrum Super Cube). FLAASH is a “Atmospheric Correction” software package based on MODTRAN, developed by AirForcePhii1ips Laboratoiy and SpectraiSciences, Inc. of Hanscom Air Force Base. The software supports current and planned infrared-visible light-ultraviolet hyperspectral and multispectral sensors. The software provides accurate physical derivation of surface and atmospheric characteristics (such as ground π albedo, surface surface degree, water vapor, optical depth of aerosols and clouds, surface and temperature). (D) Receive the spectral characteristics of the material of interest from the spectral library 50, and pre-process the characteristics in the characteristic pre-processing module 4Q. Spectrum library is a database of known materials and objects, and provides its special 83 · 3.10,000 (please read the precautions on the back before filling this page) Order -10- 503375 A7 B7 V. Description of the invention (9 Levy 10 15 , Itchyhoof # 4¾¾¾¾ 20 Some of the features obtained from the laboratory are the reflectance corresponding to the relative wavelength of a single material. The features from the spectral library are orthogonally normalized in the feature pre-processing module 40. This process decomposes the feature space Into a set of orthogonal vector sets. They are linear combinations of known materials or objects. 5 (e) The first pixel content estimation module 90 receives the orthogonal features from the feature pre-processing module 40 and the atmospheric features from the atmospheric compensation module 30. The first pixel is measured and the content vector of the first pixel is calculated. The content estimate of the first pixel will be used as the initial value of the robust filter 60. For pixels other than the first pixel, the first The pixel content estimation module 90 bypasses the measurement data from the atmospheric compensation module 30. There are different algorithms for estimating the content vector of the first pixel. The available algorithms include the minimum Multiplicative (LS) estimator, maximum likelihood (ML) 'and evolutionary algorithm (EA). (F) The robust filter receives orthogonal features from the feature pre-processing module 40 and current pixels from the atmospheric compensation module 30 Point measurement data, and content estimation from the content vector estimation module 90 of the first pixel, and use Kalman filtering technology to perform spectral decomposition of the current pixel. The robust filter 60 outputs the content estimation of the current pixel to the genetic algorithm The decomposition module 70. (g) The genetic algorithm decomposition module 70 receives the orthogonal features of the feature pre-processing module 40, the current pixel measurement value from the atmospheric compensation module 30, and the current pixel content estimation from the robust filter 60. , And use genetic algorithm to perform accurate spectral decomposition of the current pixel to obtain an accurate content estimate. -11- ---------- (Please read the precautions on the back before filling this page)-Order this paper The scale is applicable to the Chinese National Standard (CNS) A4 (210X297 mm) 83 · 3 · 丨 〇, 00〇503503 A7 B7 V. Description of the invention (10) The accurate content estimate is the system output. At the same time, The value is also fed back to the robust filter 60 in the robust filter as a new initial value for the content estimation of the next pixel. In this step, the output of the robust filter 60 is used as a genetic algorithm decomposition 5 module 7. Starting point to accelerate the northbound pass of the genetic algorithm / ·-, (h) move to the next pixel point, and the robust filter 60 receives the orthogonal features from the special feature pre-processing module 40 from the atmospheric compensation The current pixel measurement value of module 30 and the content estimation of the previous pixel from the genetic decomposition module 70, and the Kalman filtering technology is used to perform frequency 10 spectral decomposition on the current pixel. In this step, the genetic algorithm decomposition The accurate content estimation of module 70 is used in the robust filter 60 as the content estimation of the previous pixel in order to accurately estimate the content vector of the current pixel. (i) Return to step (g), and repeat steps (g) and (h) in the hyperspectral characteristic cube until the last pixel. The value of 15 points out that the robust filter 60 and the genetic algorithm can run in parallel. As shown in the fourth figure, after step (d), another step is added, that is, the artificial neural network module 80 is used for material classification selection. This step can reduce the elements of the content vector to accelerate the analysis of hyperspectral characteristics. Therefore, after step (d), the method of the present invention consists of the following steps. (e ') The artificial neural network 80 receives the orthogonal features from the feature pre-processing module 40 and the current pixel measurement data from the atmospheric compensation module 30, and performs material classification selection. The classification data is sent to a genetic algorithm decomposition module 70 and a robust filter 60. The classification data of the first pixel point is also sent to the first image. -12- This paper Caiguan County (CNS) A4 · (21GX297 public director) 83.3.10,000 (Please read the precautions on the back before filling this page) Order 5 Description of the invention (ll A7 B7 J point content purchase 0. _, artificial nerve _ she bypasses the measurement data from the atmospheric compensation module 30 to subsequent modules. ) Is called = class, and is used by human I neural network _ in correction sensor _. The corrected pixel input data from the current compensation module 30 is processed by artificial neural network to record the reading sign and face the known difficulty. The correlation degree of ^. ★ In the fuzzy neural network, the output of all neurons is used to evaluate to determine whether any neuron's response exceeds a preset value, such as · 0 · 5 (output range of neurons: 0 · 0 ~ 1 · 0) β If there are one or more neurons in the neural network whose response exceeds the preset value, the classification corresponding to the neuron is selected = continuation_candidate method. The advantage of classification selection is that it is only related to the current pixel. Related material content is only in Kalman filter and genetic calculation It is used as the element of the state vector in this method. This step can reduce the number of elements of the content vector, 15 to speed up the process. (F ') The first pixel content estimation module 90 receives the JE intersection features from the feature preprocessing module, Compensation module 3_ The first pixel measurement data and the classification data of the first pixel from the artificial neural network 80, and calculate the content vector of the first pixel. The content estimation value of the first pixel is used as The initial value of the robust filter 60. For pixels other than the first pixel, the first pixel content estimation module 90 will only bypass the measurement data from the artificial neural network 80. There are different algorithms available for the first The algorithm of content vector estimation of pixels includes the least squares (LS) economizer and the maximum likelihood method.

(ML I (請先閲讀背面之注意事項再填寫本頁)(ML I (Please read the notes on the back before filling this page)

、1T -13- 冗張从適用中國2iOX297公瘦) 83. 3· 1〇,〇〇〇 503375 A7 B7 五、發明説明(12) ),及進化算法(EA)。 (g’)魯棒濾波器6〇接收特徵預處理模塊4〇的正交特 徵’大氣補償模塊3〇的當前像素點的測量數據,人工神經 網絡80的分類數據,第一像素點含量估計模塊90的第一像 5素點含量估計,並利用卡爾曼濾波技術對當前像素點進行 頻譜兮解。魯棒濾波器60输出當前像素點的含量估計給遺 傳算法分解模塊7Q。 (h’)遺傳算法分解模塊70接收特徵預處理模塊40的 正交特徵,大氣補償模塊30的當前像素點的測量數據,人 10工神經網絡80的分類數據,第一像素點含量估計模塊9〇的 第一像素點含量估計,並利用遺傳算法對當前像素點進行 ‘確頻譜分解從TO得出準確含量估計。該準確含量估計值 即為系統输出。同時,該準確含量估計值也反饋到魯棒濾 波器60,在魯棒濾波器中用作為下一個像素點含量估計的 15新初始值。在這一步,魯棒濾波器6〇的输出被用作為遺傳 算法分解模塊的起始點,以便加速遺傳算法的進化。 (Γ )移到下一個像素點,魯棒濾波器6〇接收來自特 徵預處理模塊40的正交特徵,來自大氣補償模塊3〇的當前 像素點測量值,人工神經網絡80的分類,以及來自遺傳分 20解模塊70的前一個像素點的含量估計,並利用卡爾曼濾波 技術對當前像素點進行頻譜分解。在這一步中,來自遺傳 算法分解模塊70的準確含量估計在魯棒濾波器go中用作前 一個像素點的含量估計以便準確估計當前像素點的含量矢 量。 -14- 本紙張尺度適用中國國家梂準(CNS ) A4規格(210X297公釐) 83. 3.10,000 I--------m— C请先閱讀背希之注意事項存填寫本 503375 A7 B7 五、發明说明(13 ) ㈠’)回到步驟(h,),在高譜特性立方體內重復遞推 步驟(h’)與(Γ )直到最後-憾素點。 (請先閎讀背面之注意事項再填寫本頁} 在某些應用中,處理速度最為重要,而含量估計準確 性則是次要的,那麼,步驟(g)可由以下過程替代: 5 評估當前像素點的含量估計誤差。如果估計誤差大於 預設值,那麼執行步驟(g),其中估計誤差由魯棒濾波器 、的協方差矩陣給出。否則,遺傳算法分解模塊70僅僅繞過 魯棒濾波器60的解。 類似地,步驟(h,)可由以下過程替代: 10 評估當前像素點的含量矢量估計誤差。如果該估計誤 差大於預設值,那麼執行步驟(h,),其中估計誤差由魯棒 濾波器的協方差矩陣給出。否則,遺傳算法分解模塊7〇僅 僅繞過魯棒濾波器的解。 參考第七圖,遺傳算法分解模塊7〇的遺傳算法分解過 15程進一步由以下步驟組成: (1) 編碼模塊76隨機生成而一組二進制字符串π。 這組一進制字符串代表高譜圖象立方體的一個像素點的含 量矢量。遺傳算法對這組二進制字符串進行操作,而不是 對含量矢量本身進行操作β二進制字符串送到解碼模塊72 20 ° (2) 解碼模塊72對這組二進制字符串77進行解碼。 解碼模塊的输出是關於超級頻譜圖象立方體像素點的含量 矢量集合。這級含量矢量集合被送到適合度計算模現。 含量矢量給出了該像素點中所包含的每個感興趣材料的百 -15- ( CNS ) A4^( 210X297^* )-~---^_ 83. 3.10,000 503375 A7 -------- B7 五、發明説明(χ4) " ^— --— 分比。 (3) 適合度雜麵71計算雜含量矢量的適合度 值、。在適合度什算模塊中,指標函數取一個二進制字符串 並返回-個值。該一進制字符串也稱之為染色體。然後 5,指標函數的值映射為適合度以適應遺傳算法。適合^值 是基於由該字符串代表的所有可能解的性能的回答。編瑪 字符串(染色體)的解(含量矢量)越好,適合度值越高 。適合度值再送到繁殖模塊73 α (4) 在繁殖模塊73中,基於來自適合度計算模塊71 W的適合度输出,進行繁殖。繁殖是基於適者生存的規律。 這些具有高適合度值的字符串在新一代中被大量複製。一 旦這些字符串被繁殖或複製作為下-代使用,它們將在配 對集合中進行另外兩類操作,即交配與變異,從而繁殖。 (5) 在交配模塊74中,通過交換字符串(染色體)的 15頭和尾來形成子字符串集合。交配為二進制字符串提供了 一種機制通過隨機過程來混合和匹配其期望的品質。首先 ’從繁殖模塊73形成的匹配集合中選出兩個新生成的字符 串。其二,沿這兩個字符串一致地隨機選擇一個交换位置 β其三,交换該交换位置以後的所有字符。儘管交配使^ 2〇隨機選擇,但不能將此看作是在搜尋空間隨機漫步。當與 繁殖過程相結合時,這是一種交換信息並形成高質量解的 有效手段。 • (6)在變異模塊75中,偶而改變在一個特定字符:串 位置上的值。這個步驟增強遺傳算法找到一個近似最優解 -16- 本纸浪尺度逋用中國國家揉準(CNS ) Μ規格(210X297公釐) 83· 3. !〇,〇〇〇 ^-- (請先閲讀背面之注意事項再填寫本頁) 訂 五、發明説明(一-~----- 的能力僅異是肺何解位_久損賴―魏險措施 °變異的發生概率極低,以至在字符串集合中平^一 個字符串發生變異。 ^ 、(7)將新的二進制字符串集合送到解碼模塊72,然 5後按(2),(3),⑷,⑸,⑹,以及⑺步驟執行^ 在步驟(3)中,執行甄別過程以確定是否終止進化。 、甄別指標定義為總進化代數。當遺傳算法遞推到總進化代 數時選擇個具有取大適合度值的二進制字符串作為解 。其相應的含1:矢1:就是諭素賴含量料。也可通過 10評價字符串間的差別來進行甄別。如果字符串間的差別小 於-個預設值,則退出進化。同樣,選擇_個具有最大適 合度值的二進制字符串作為解。 因為卡爾曼濾波器根據定義好的統計特性產生最優估 計,估計值是無偏差的,且在線性無偏差估計^具有最小 15方差。然而,其估計質量僅僅在數學模型成立的假設下得 到保證β模型中的任何誤差都可能使濾波結果無效,因而 ’基於該結果的任何結論同樣都無效β 在本發明的混合方法中’對於咼譜分解,魯棒濾波器 的另一個選擇是萬能魯棒卡爾曼濾波器。該萬能魯棒卡爾 20曼濾波器能在多個動態環境/系統下穩定運行。 參考第五圖,高譜分解的萬能魯棒濾波方法由以下步 驟組成·· (f.l)接收來自大氣補償模塊30的像素點測量數據 -17- 本纸張尺度適用中國國家梂準(CNS ) Α4規格(210X297公釐) 8 3· 3. !0,〇〇〇 --------—-I (請先閲讀背面之注意事項再填寫本頁) 訂 發明説明(16 15 201T -13- redundant (applicable to China 2iOX297 public thin) 83.3.10,000,000 503375 A7 B7 V. Description of the invention (12)), and evolutionary algorithm (EA). (g ') The robust filter 60 receives the orthogonal features of the feature pre-processing module 40, the measurement data of the current pixels of the atmospheric compensation module 30, the classification data of the artificial neural network 80, and the first pixel content estimation module. The content of the first pixel of 90 is estimated at 5 pixels, and the Kalman filtering technique is used to perform spectral analysis on the current pixel. The robust filter 60 outputs the content estimation of the current pixel to the genetic algorithm decomposition module 7Q. (h ') The genetic algorithm decomposition module 70 receives the orthogonal features of the feature pre-processing module 40, the measurement data of the current pixels of the atmospheric compensation module 30, the classification data of the artificial neural network 80, and the first pixel content estimation module 9 The content of the first pixel point of 〇 is estimated, and the genetic pixel is used to perform a true spectral decomposition on the current pixel point to obtain an accurate content estimate from TO. This accurate content estimate is the system output. At the same time, the accurate content estimation value is also fed back to the robust filter 60, which is used as a new initial value for the next pixel content estimation in the robust filter. In this step, the output of the robust filter 60 is used as the starting point of the genetic algorithm decomposition module in order to accelerate the evolution of the genetic algorithm. (Γ) moves to the next pixel, the robust filter 60 receives the orthogonal features from the feature pre-processing module 40, the current pixel measurement from the atmospheric compensation module 30, the classification of the artificial neural network 80, and The genetic pixel 20 solution module 70 estimates the content of the previous pixel, and uses Kalman filtering technology to perform spectral decomposition on the current pixel. In this step, the accurate content estimation from the genetic algorithm decomposition module 70 is used in the robust filter go as the content estimation of the previous pixel in order to accurately estimate the content vector of the current pixel. -14- The size of this paper is applicable to China National Standard (CNS) A4 (210X297 mm) 83. 3.10,000 I -------- m— C Please read the precautions for your memory and fill in this 503375 A7 B7 5. Description of the invention (13) ㈠ ') Go back to step (h,), and repeat the recursive steps (h') and (Γ) in the hyperspectral characteristic cube until the final -regret point. (Please read the notes on the back before filling this page} In some applications, the processing speed is the most important, and the content estimation accuracy is secondary. Then, step (g) can be replaced by the following process: 5 Evaluate the current Pixel content content estimation error. If the estimation error is greater than a preset value, step (g) is performed, where the estimation error is given by a robust filter and a covariance matrix. Otherwise, the genetic algorithm decomposition module 70 only bypasses the robustness Solution of filter 60. Similarly, step (h,) can be replaced by the following process: 10 Evaluate the content vector estimation error of the current pixel. If the estimation error is greater than a preset value, then step (h,) is performed, where the estimation error It is given by the covariance matrix of the robust filter. Otherwise, the genetic algorithm decomposition module 70 only bypasses the solution of the robust filter. With reference to the seventh figure, the genetic algorithm decomposition of the genetic algorithm decomposition module 70 is further divided by 15 The following steps are composed: (1) The encoding module 76 randomly generates a set of binary character strings π. This set of unary character strings represents a pixel of a hyperspectral image cube. The quantity vector. The genetic algorithm operates on this set of binary strings instead of the content vector itself. The β binary string is sent to the decoding module 72 20 ° (2) The decoding module 72 decodes this set of binary strings 77. Decode The output of the module is a set of content vectors about the pixels of the super-spectral image cube. This set of content vectors is sent to the fitness calculation model. The content vector gives the percentage of each material of interest contained in the pixel. -15- (CNS) A4 ^ (210X297 ^ *)-~ --- ^ _ 83. 3.10,000 503375 A7 -------- B7 V. Description of Invention (χ4) " ^---- (3) The fitness degree miscellaneous surface 71 calculates the fitness value of the miscellaneous content vector. In the fitness degree calculation module, the index function takes a binary string and returns a value. The unsigned string is also called It is a chromosome. Then, the value of the indicator function is mapped to a fitness to adapt to the genetic algorithm. The fit value is an answer based on the performance of all possible solutions represented by the string. The solution (content) of the string (chromosome) is compiled. Vector) The better, the fitness value High. The fitness value is then sent to the breeding module 73 α (4) In the breeding module 73, breeding is performed based on the fitness output from the fitness calculation module 71 W. Reproduction is based on the law of survival of the fittest. These have high fitness values The strings are reproduced in large numbers in the new generation. Once these strings are propagated or copied for use as the next generation, they will perform two other types of operations in the pairing set, namely mating and mutation, thereby multiplying. (5) In mating In module 74, a set of substrings is formed by exchanging the 15 heads and tails of a string (chromosome). Mating provides a mechanism for binary strings to mix and match their desired qualities through a random process. First, 'two newly generated character strings are selected from the matching set formed by the breeding module 73. Second, randomly select a swap position β along the two character strings. Third, swap all characters after the swap position. Although mating makes ^ 20 randomly selected, this cannot be seen as a random walk in the search space. When combined with the reproduction process, this is an effective means of exchanging information and forming high-quality solutions. • (6) In the mutation module 75, the value at a specific character: string position is occasionally changed. This step enhances the genetic algorithm to find an approximate optimal solution. -16- The scale of this paper uses the Chinese National Standard (CNS) M specification (210X297 mm) 83 · 3.! 〇, 〇〇〇 ^-(please first Read the precautions on the back and fill in this page) Order five. Invention description (a- ~ ----- The ability is only different from the position of the lungs _ long-term loss ─ Wei Wei measures ° The probability of occurrence of mutation is extremely low, A string is mutated in the string set. ^, (7) Send the new binary string set to the decoding module 72, and then press (2), (3), ⑷, ⑸, ⑹, and执行 Step execution ^ In step (3), a screening process is performed to determine whether to terminate the evolution. The screening index is defined as the total evolution algebra. When the genetic algorithm is recursive to the total evolution algebra, a binary character with a large fitness value is selected. The string is used as a solution. The corresponding content is 1: vector 1: is the content of 谕 Sulai content. It can also be identified by evaluating the difference between 10 strings. If the difference between the strings is less than a preset value, exit evolution. Similarly , Select _ binary characters with the largest fitness value As a solution. Because the Kalman filter produces an optimal estimate based on the defined statistical characteristics, the estimate is unbiased and has a minimum of 15 variance in the linear unbiased estimate ^. However, its estimation quality is only based on the assumption that the mathematical model holds. It is guaranteed that any error in the β model may invalidate the filtering result, so 'any conclusion based on this result is also invalid β in the hybrid method of the present invention' Another option for a robust filter for unitary spectral decomposition is Universal robust Kalman filter. The universal robust Kalman filter can run stably under multiple dynamic environments / systems. Referring to the fifth figure, the universal robust filtering method for hyperspectral decomposition consists of the following steps ... ( fl) Receive pixel measurement data from the atmospheric compensation module 30-17- This paper size applies to China National Standard (CNS) Α4 specification (210X297 mm) 8 3 · 3.! 0, 〇〇〇〇 ---- ----—- I (Please read the notes on the back before filling this page) Order the invention description (16 15 20

I (f· 2 )在模糊邏輯模塊61中,利用模糊邏辑推理方 法檢驗測量數據。在此,模糊邏辑推理方法依據模糊邏辑 規則確定拒絕或校正測量數據,或承認該測量數據。 (f.3)输出正確的測量數據或錯誤標誌到預處理模 塊62,在此,預處理模塊62執行狀態轉移矩陣和測量矩陣 的計算。 (f· 4)將計算好的狀態轉移矩障從預處理模塊62送 到狀態矢量預測模塊66,將前一個狀態矢量從狀態矢量更 新模塊68送到狀態矢量預測模塊66,在此,狀態矢量預測 模塊66進行狀態矢量預測,也就是下一個像素點的含量矢 量。 (f. 5)將計算好的狀態轉移矩障從預處理模塊似送 到協方差傳播模塊63,在此,協方差傳播模塊63計算當前 估計誤差的協方差。 (f· 6)將測量矩陣及當前測量矢量從預處理模塊62 送·到測量殘差計算模塊67,在此,測量殘差計算模塊67接 收來自狀態矢量預測模塊66的狀態矢量預測值,並通過從 當前測量矢量中減去測量矩陣與狀態矢量預測值的乘積來 計算測量殘差。 (f· 10將當前估計誤差的協方差從協方差傳播模塊 63送到最優增益計算模塊64,在此,最優增益計算模塊64 執行最優增益計算。 (f. 8)將最優增益從最優增益計算模塊64送到協方 差更新模塊65,在此,協方差更新模塊64更新估計誤差的 *18- 本紙張纽ϋ用巾us家;^ (CNS)〜胁(21QX297公羡〉 83. 3.10,000 (請先閲讀背面之注意事項再填寫本頁) -^1. 訂 線- 503375 A7 B7 五、發明説明(17) 10 協方差。 (f· 9)將估計誤差的協方差更新值從協方差更新模塊 65送到協方差傳播模塊63。 (f· 10)將最優增益從最優增益計算模塊64送到狀態 =量更新魏68,在此,狀態矢量更新模塊嶋收來自測 里殘差计算模塊66的測量殘差,並進行狀態矢量更新,也 即,下一個像素點的含量矢量。 參考第六圖,步驟(f· 2)進一步由以下步驟組成: ^ (f·2—1)將測量數據送到模糊化模塊611,在此,模 糊器執行標量映射,也就是把測量範圍轉換到一個相應的 論域,並進行模糊化,也就是把·數賴換成適當的語 義值。這些語義值被標為模糊集合,並對所得的模糊輸入 用模糊集合及其隸屬函數(〔〇, 1〕)解釋確定的測量數據 (請先閲讀背面之注意事項再填寫本頁) 訂 15 讀崔 (f· 2-2)將模糊翰入從模糊器模塊611送到模糊推理 機構612。在此,模糊推理機構612本質上模仿人類的決策 機理,並利用模糊推理規則推出模糊输出。 來自模糊規則庫613的模糊邏輯推理規則借助一組語 義規則歸納了專家的目標和策略。模糊規則庫由應用域知 2〇識及目標組成。 (L2-3)將模糊输出從模糊推理機構612送到模糊還 原模塊614,在此,模糊還原模塊614生成一個確定的測量 數據,該數據最好地表達了推理模糊输出的可能分布。 -19- 本紙張尺度適用中國國家榡準(€阳)六4^^(210><297公釐) 83. 3. 10,000 線一I (f · 2) In the fuzzy logic module 61, the measurement data is checked by the fuzzy logic inference method. Here, the fuzzy logic inference method determines whether to reject or correct the measurement data or to acknowledge the measurement data according to the rules of the fuzzy logic. (f.3) Output correct measurement data or error flags to the pre-processing module 62, where the pre-processing module 62 performs calculation of the state transition matrix and the measurement matrix. (f · 4) The calculated state transition moment barrier is sent from the preprocessing module 62 to the state vector prediction module 66, and the previous state vector is sent from the state vector update module 68 to the state vector prediction module 66. Here, the state vector The prediction module 66 performs state vector prediction, that is, the content vector of the next pixel. (f. 5) The calculated state transition moment barrier is sent from the preprocessing module to the covariance propagation module 63. Here, the covariance propagation module 63 calculates the covariance of the current estimation error. (f · 6) The measurement matrix and the current measurement vector are sent from the preprocessing module 62 to the measurement residual calculation module 67. Here, the measurement residual calculation module 67 receives the state vector prediction value from the state vector prediction module 66, and The measurement residual is calculated by subtracting the product of the measurement matrix and the predicted value of the state vector from the current measurement vector. (f · 10 sends the covariance of the current estimation error from the covariance propagation module 63 to the optimal gain calculation module 64, where the optimal gain calculation module 64 performs the optimal gain calculation. (f. 8) The optimal gain The optimal gain calculation module 64 is sent to the covariance update module 65, where the covariance update module 64 updates the estimated error * 18- This paper is used as a paper towel; ^ (CNS) ~ Waki (21QX297 public envy> 83. 3.10,000 (Please read the notes on the back before filling this page)-^ 1. Ordering line-503375 A7 B7 V. Description of the invention (17) 10 Covariance. (F · 9) Covariance of estimated error The updated value is sent from the covariance update module 65 to the covariance propagation module 63. (f · 10) The optimal gain is sent from the optimal gain calculation module 64 to the state = quantity update Wei 68, where the state vector update module receives The measurement residual from the measurement residual calculation module 66 is updated and the state vector is updated, that is, the content vector of the next pixel. Referring to the sixth figure, step (f · 2) is further composed of the following steps: ^ (f 2-1) Send the measurement data to the fuzzification module 611, where the fuzzer executes Quantitative mapping, that is, transforming the measurement range to a corresponding domain and blurring it, that is, changing the number data to appropriate semantic values. These semantic values are labeled as fuzzy sets, and the resulting fuzzy inputs are used. The fuzzy set and its membership function (〔〇, 1〕) explain the determined measurement data (please read the notes on the back before filling this page) Order 15 Reading Cui (f · 2-2) Put the fuzzy input into the fuzzer module 611 is sent to the fuzzy inference mechanism 612. Here, the fuzzy inference mechanism 612 essentially imitates human decision-making mechanism and uses fuzzy inference rules to derive fuzzy outputs. The fuzzy logic inference rules from the fuzzy rule base 613 summarize the experts with a set of semantic rules Goals and strategies. The fuzzy rule base consists of application domain knowledge and goals. (L2-3) The fuzzy output is sent from the fuzzy inference mechanism 612 to the fuzzy reduction module 614. Here, the fuzzy reduction module 614 generates a certain Measured data, this data best represents the possible distribution of inferred fuzzy output. -19- This paper scale applies to China National Standards (€ yang) 6 4 ^^ (210 > < 297 PCT) 3. 83. Line 10,000

Claims (1)

503375 A8 B8 C8 D8 申請專利範圍 15 經濟部智慧財產局員工消費合作社印製 ι·-高譜圖象分析及像素點頻譜分解方法,由以下步驟板 成: 、(a)接收來自-高譜圖象傳感器的—個高譜圖象立 ^體,該高譜圖象立方體以波長及空間位置表示一個場 景, .(b)在-®象登記模财,進行逐個波段圖象登記, 該圖象登記是使-個波段的圖象轉換到另—個波段圖象而 不必涉及細賴魏,轉證對應在—個波段圖象中 的一個像素點的物理位置與在另一個波段圖象中的像素點 的物理位置一致; t (〇發送在該圖象登記模塊中登記後的高譜圖象立 方體到大氣補償模塊,在該大氣補償模塊内補償大氣效 應;及 (d)利用遺傳算法與魯棒卡爾曼濾波技術以逐個像 素點方式分解南譜圖象數據,直到完成該高譜圖象立方體 者。 2·如申凊專利範圍第1項所述之「高譜圖象分析及像素點頻 譜分解方法」,其中在步驟(c)中,大氣校正可用上市 軟體FLAASH (頻譜超級立方體的快速視線大氣分析)來 完成,該FLAASH是一個基於MODTRAN的“大氣校正” 軟體包’由Hanscom 空軍基地的 AirForcePhillipsLaboratory 與SpectralSciencesJnc.開發,該軟體支持當前及規劃的紅 外線-可見光-紫外線高譜及多頻譜傳感器,該軟體提供 了地表及大氣特性(諸如,地表反照率,地表高度,水蒸 -20- 本纸張^度逋用中國國家標準(CNS ) A^i格(210Χ297公釐) 請 先 閱 讀 背 面 之 注 意 事 項 再 填 寫 本 頁 裝 π 線 503375 A8 B8 C8 D8503375 A8 B8 C8 D8 Application for patent scope 15 Printed by the Consumer Cooperatives of Intellectual Property Bureau of the Ministry of Economic Affairs-High-spectrum image analysis and pixel spectral decomposition method, composed of the following steps: (a) Receive from-Hyperspectral A hyperspectral image cube like a sensor, the hyperspectral image cube represents a scene by wavelength and spatial position. (B) Register image by band in the -® image registration mode. This image The registration is to convert one band image to another band image without involving Lai Weiwei. The transfer certificate corresponds to the physical position of a pixel in the one band image and the physical position of the pixel in the other band image. The physical locations of the pixels are consistent; t (0) sends the hyperspectral image cube registered in the image registration module to the atmospheric compensation module, and compensates for atmospheric effects in the atmospheric compensation module; and (d) using genetic algorithms and Lu The rod Kalman filtering technology decomposes the South-spectrum image data on a pixel-by-pixel basis until the hyperspectral image cube is completed. 2. "Hyperspectral image analysis and pixel spectrum as described in item 1 of the patent application "Solution method", in step (c), atmospheric correction can be completed with the listed software FLAASH (Fast Sight Atmospheric Analysis of Spectral Supercube), a FLATASH based "MODIFICATION" atmospheric correction "software package from Hanscom Air Force Base Developed by AirForcePhillipsLaboratory and SpectralSciencesJnc., The software supports current and planned infrared-visible-ultraviolet hyperspectral and multispectral sensors. The software provides surface and atmospheric characteristics (such as surface albedo, surface height, water vapor-20-paper) Zhang ^ Degree uses Chinese National Standard (CNS) A ^ i (210 × 297 mm) Please read the precautions on the back before filling in this page. Π Line 503375 A8 B8 C8 D8 申請專利範圍 15 經濟部智慧財產局員工消費合作社印製 20 汽,浮質與雲層的光學深度,地表與溫度)的準確物 推導者。 3.如申請專纖圍第丨項所述之「高_象分析及像素點頻 請分解方法」,其中,步驟⑷進—步細下步雜構成: (d-Ι)接收來自頻譜庫的感興趣材料的頻譜特徵, 並在特徵預處理模塊預處理這些特徵,來自該頻^譜庫的這 些頻譜特徵在該特徵預處理模塊令進行正交歸一化,從而 將該特徵空間分解成一組正交特徵集合; (d 2)第一像素點含量估計模塊接收來自該特徵預 處理模塊的這些正交特㈣絲自該缝補償模塊的一組 第一像素點測量數據,並計算該第一像素點的含量矢量’ 除該第一像素點以外的像素點,該第一像素點含量估計模 塊僅僅繞過來自該大氣補償模塊的這些測量數據; 、 (d-3)卡爾曼濾波器接收來自該特徵預處理模塊的 這些正交特徵,來自該大氣補償模塊的當前像素點測量數 據,以及該第一像素點含量矢量估計模塊的含量估計:並 用卡爾曼濾波技術進行當前像素點的頻譜分解,該第一像 素點的含量估計用作為該卡爾曼濾波器的初始值,該卡爾 曼遽波器輸出該當前像素點的含量估計; (d~4)遣傳算法分解模塊接收該特徵預處理模塊的 正交特徵,來自該大氣補償模塊的該當前像素點測量值, 以及來自該卡爾曼濾波器的當前像素點的含量估計,並利 用該遺傳算法對該當前像素點進行準確頻譜分解以得出準 確含量估計; (請先閎讀背面之注意事項再填寫本 -裝 訂 線· -21-Scope of patent application 15 Printed by the Consumer Cooperatives of the Intellectual Property Bureau of the Ministry of Economic Affairs 20 The accurate derivation of vapor, optical depth of aerosols and clouds, surface and temperature). 3. As described in the application of "Special Fiber Circumstances", the "High-Analysis and Pixel Frequency Decomposition Method", in which, step by step-step by step and step by step: (d-1) Receive the The spectral characteristics of the material of interest, and these characteristics are pre-processed in the feature pre-processing module. The spectral features from the frequency spectrum library are orthogonally normalized in the feature pre-processing module, thereby decomposing the feature space into a group Orthogonal feature set; (d 2) The first pixel content estimation module receives a set of first pixel measurement data of the orthogonal special filaments from the feature preprocessing module from the slit compensation module, and calculates the first The content vector of the pixels' pixels other than the first pixel, the first pixel content estimation module only bypasses these measurement data from the atmospheric compensation module; (d-3) the Kalman filter receives from The orthogonal features of the feature preprocessing module, the current pixel measurement data from the atmospheric compensation module, and the content estimation of the first pixel content vector estimation module: and use the Kalman filtering technique Perform spectral decomposition of the current pixel, the content estimate of the first pixel is used as the initial value of the Kalman filter, and the Kalman waver outputs the content estimate of the current pixel; (d ~ 4) the retransmission algorithm The decomposition module receives the orthogonal features of the feature preprocessing module, the current pixel measurement value from the atmospheric compensation module, and the current pixel content estimation from the Kalman filter, and uses the genetic algorithm to the current pixel Points to perform accurate spectral decomposition to obtain accurate content estimates; (Please read the notes on the back before filling in this-gutter · -21- 、申請專利範圍 10 15 經濟部智慧財產局員工消費合作社印製 20 (心5)移到下一個像素點,該卡爾曼濾波器接收來 =特徵猶理模塊的正交特徵,來自該大氣補償模塊的 :則像素綱餘,収來自輯傳分賴塊的前一個像 :、點的合I估計,並利用卡爾曼濾波技術對該當前像素點 進行頻譜分解者。 1 如X申請專利範圍第3項所述之「高譜圖象分析及像素點頻 解方法」’其中在步驟⑷)_,該科含量估計 疋不,輪出,且該當前像素點的準確含量估計被反饋給該 卡爾叉瀘波器,在該卡爾曼濾波器中,用作為下一個像素 點的個含量估計的初始值,該卡爾曼濾波器的輸出用 遺傳算法分解模塊的起始點以便加速遺傳算法的進 申請專利範圍第4項所述之「高譜圖象分析及像素點頻 j分解方法」,其中在步驟(d-5)中,該遺傳算法分解 柄塊的該準確含量估計在該卡岐渡波器中用作為該前一 個像素點的含量估計以便準確估計該當前像素點的含量矢 量者。 穴 申請專利範圍第5項所述之「高譜圖象分析及像素點頻 ^曰分解方法」,其中在步驟(d-2)中,計算該第一像素 點的含量估計的參數估計器是一個最小二乘(LS)估計器 者。 ^如、申清專利範圍第5項所述之「高譜圖象分析及像素點頻 %刀解方法」,其中在步驟(d-2)中,計算該第一像素 點的含量估計的參數估計器是一個極大似然(ML)估計器 -22- 本紙張劇中國 ---—-----^ηϋτ. (請先閎讀背面之注意事項再填寫本頁) 、aT 線 1......1 I -----I »*---1 ·- -II · 六、申請專利範圍 5 10 15 8 8 8 8 ABCD 經濟部智慧財產局員工消費合作社印製 20 S·如申請專鄕圍第5項所述之「 级八銥十土 °曰圖象分析及像素點$ °曰刀解方法」,其中在步驟⑷)中,計嘗該第 點的含量估計的參絲龍是—個進化算^者。i 9_如申請專利範圍第丨項所 圖^ 請分解方法」,料轉⑷由及像素制=1)触來自觸庫的感興趣材料的頻譜 站特徵聽職塊預處輯麵徵, 些頻譜舰在賴徵減職财進行正线—化庫^ 將該特徵空間分解成一組正交特徵集合; 一⑷)人I神經·触來自該特徵親理模塊的 ^正父特徵以及來自該大氣補償模塊的—组第一像素黑 測據,並進行㈣分_擇,該人碎經網絡繞過拉 該大氣補償模塊到後續模塊的這些測量數據; (do)第一像素點含量估計模塊接收來自該特徵預 處理模塊的這些正交舰,來自該大氣補償模塊的一級第 -像素點測量數據,以及來自該人工神經網絡的苐一像素 點的分類數據,並彻-個參紐計輯算該第—像素點 ^含量矢量,雜-像素關含量料用.作為甸曼濾波 器的初始值,除該第一像素點以外的像素點,該第一像素 點含2:估計模塊僅僅繞過來自該人工神經網絡模塊的這也 測量數據; 〜 (孓4)該卡爾曼濾波器接收來自該特徵預處理模塊 的追些正交特徵,來自該大氣補償模塊的當前像素點測量 -23- 本纸張又度適用t國國家#準(CNS ) A4規格(21〇Χ297公釐) (請先閎讀背面之注意事項再填寫本頁} : 、-*&amp;Scope of patent application 10 15 Printed by the Intellectual Property Bureau of the Ministry of Economic Affairs's Consumer Cooperatives 20 (Heart 5) Moved to the next pixel, the Kalman filter receives the orthogonal feature of the = feature module, from the atmospheric compensation module : Then the pixel outlines are received from the previous image of the pass-by block: the combined I estimation of the points, and the current pixel point is spectrally decomposed by using Kalman filtering technology. 1 As described in item 3 of the scope of X's patent application, "Hyperspectral image analysis and pixel frequency solution method", where in step ⑷) _, the content of this section is estimated to be inactive, and the current pixel is accurate The content estimate is fed back to the Kalman waver. In the Kalman filter, the initial value of the content estimation of the next pixel is used. The output of the Kalman filter is decomposed by the starting point of the genetic algorithm module. In order to speed up the genetic algorithm's application of the "Hyperspectral Image Analysis and Pixel Frequency j Decomposition Method" described in item 4 of the patent scope, in step (d-5), the genetic algorithm decomposes the accurate content of the handle block. The estimation is used as the content estimation of the previous pixel point in the Kaqi Tow to accurately estimate the content vector of the current pixel point. The "Hyperspectral image analysis and pixel frequency decomposition method" described in item 5 of the patent application scope, wherein in step (d-2), the parameter estimator for calculating the content estimation of the first pixel is A least squares (LS) estimator. ^ The "Hyperspectral image analysis and pixel point frequency% knife solution method" as described in item 5 of the scope of patent application, wherein in step (d-2), parameters for estimating the content of the first pixel point are calculated The estimator is a Maximum Likelihood (ML) Estimator. -22- This paper drama China --------------- ^ ηϋτ. (Please read the notes on the back before filling this page), aT line 1. ..... 1 I ----- I »* --- 1 ·--II · VI. Application scope 5 10 15 8 8 8 8 ABCD Printed by the Consumer Cooperatives of the Intellectual Property Bureau of the Ministry of Economic Affairs 20 S · For example, the "level eight iridium and ten soil ° image analysis and pixel point resolution method" described in item 5 of the application, wherein in step ii), the parameters of the content estimation of this point are considered. Silk Dragon is an evolutionary operator. i 9_As shown in item 丨 of the scope of patent application ^ Please decompose the method ", the material conversion and pixel system = 1) touch the spectral station characteristics of the material of interest from the touch library. The spectrum ship performs the front-line conversion in Lai Zheng's demissions. ^ The feature space is decomposed into a set of orthogonal feature sets. 1) The human I nerve · touches the ^ positive father feature from the feature affinity module and comes from the atmosphere. Compensation module—set the first pixel black measurement data and select it. The person breaks through the network to bypass the measurement data that pulls the atmospheric compensation module to subsequent modules; (do) the first pixel content estimation module receives The orthogonal ships from the feature pre-processing module, the first-level first-pixel measurement data from the atmospheric compensation module, and the classification data of the first pixel from the artificial neural network are calculated in a single parameter. The first pixel ^ content vector is used for the miscellaneous pixel content content. As the initial value of the Dianman filter, the pixel points other than the first pixel point, the first pixel point contains 2: the estimation module only bypasses From the artificial neural network model This is also measured data of the block; ~ (孓 4) The Kalman filter receives some orthogonal features from the feature pre-processing module, and the current pixel measurement from the atmospheric compensation module -23- This paper is again applicable t 国 国 # 准 (CNS) A4 specification (21〇 × 297 mm) (Please read the precautions on the back before filling out this page}: 、-* &amp; 申請專利範圍 5 15 經濟部智慧財產局員工消費合作社印製 20 神經網絡的分類數據,以及該第-像素 進行、翻含量估計,個卡爾^皮技術 像素:^::頻譜分解,該相曼遽波器輪出該當前 正交触鱗麵處理模塊的 來白访z求自該大虱補彳員模塊的該當前像素點測量值, 經網絡的分峨據,以及來自該卡爾曼濾波 =_含量估計,並利用該遺傳算法對該當前 仃準確頻譜分解以得解確含量估計,該卡爾曼 二=itB用作為輯傳算法分解概的缺點以便加速 化,該料含量估計反制該卡爾曼遽波 二^ 用作為下__個像素點含 的新初始值; ⑷)移到下—個像素點,該卡^波器接收來 j特徵賴理模塊的歧雖,來自該域補償模塊的 田月i像素點測1值’來自該人工神經網絡的分類數據,以 及來自該遺法分解觀哺-個含量估計, 並利用卡爾叉濾波技術對該當前像素點進行頻譜分解,來 自成遺傳弃広分解模塊的準確含量估計在該卡爾曼濾波器 _用作為前-個像素_含频計,以解確估計該當前 點的含量矢量; (d-7)回到步驟(μ),在該高譜特性立方體内重復 遞推步驟(d-5)與(d-6)直到最後一個像素點者。 10·如申請專利範圍第9項所述之「高譜圖t分析及像素點 ---------襄 (請先閎讀背面之注意事項再填寫本 #1. 頁} 訂 線 -2 ‘ 本纸張逋用中國國家榡準(CNS ) ( 210x297公釐) A8 B8 C8 D8 申請專利範圍 頻°曰刀解方法」,其中在步驟(d-3)中,計算該第一像 素點的含量估計的參數估計器是-個最小二乘(LS)估計 器者。 11·如申請專利_第9項所述之「高譜圖象分析及像素點 頻办解方法」,其中在轉⑷)巾,計算該第一像 素點的s里估叶的參數估計器是一個極大似然(g)估 計器者。 12:如申請專利範圍第9項所述之「高譜圖象分析及像素點 頻4分解方法」’其令在步驟⑷)+,計算該第一像 素點的含量估計的參數估計器是-個進化算法者。 13f申請專利_第丨項所述之「高《象分析及像素點 頻°曰刀解方法」,其中步驟(d)由以下步驟構成: 15 經濟部智慧財產局員工消費合作社印製 20 ,(d 1 )接收來自頻譜庫的感興趣材料的頻譜特徵, 亚在特徵爾職塊爾理這雜徵,來自鶴譜庫的這 些頻語特徵在該特徵預處理模塊令進行正交歸—化,從而 將該特徵空間分解成一組正交特徵集合; (d 2)第一像素點含量估計模塊接收來自該特徵預 處理棋塊的這些正交特徵’以及自該大氣補償模塊的一级 第一像素點測量數據,並利芾一個參數估汁器計算該第一 像素點的含量估計,除該第—像素触外的像素點,該第 -像素點含i料概健繞縣自該錢補償模 些測量數據; (d 3)該卡爾哭/慮波器接枚來自該特徵預處理模塊 的這些正靖徵,來自該域補伽前像素點測量 -25- ΜΛ張適用中關家梂準(CNS ) A4· ( 21Qx297公幻 〜J / J申請專利範圍 15 A8 B8 C8 D8 經濟部中央標準局員工消費合作社印製 20 數據,以及該第一像素點的含量矢量估計模塊的含量估 計,並用卡爾曼濾波技術進行當前像素點的頻譜分解,該 第一像素點的含量估計用作為該卡爾曼濾波器的初始值, 該卡爾更濾波益輸出該當前像素點的含量估計到遣傳算法 分解模塊;&quot;&quot; (d-4)評價該當前像素點的含量估計值的估計誤差, &amp;該估計誤差大於預設值時,轉到步驟(心5),其中话 計誤差由卡爾曼濾波器的協方差矩陣給出,否則,轉到步 驟(d-6),該遺傳异法分解模塊僅僅繞過由卡爾曼濾波 器得到的含量估計者; (d-5)該遺傳其法分解模塊接收該特徵預處理模塊 的正交特徵,來自該大氣補償模塊的該當前像素點測量 值,以及來自該卡爾曼濾波器的當前像素點的含量估計, 並利用該遺傳其法對該當前像素點進行準確頻譜分解以得 f準峰含量估計,該準確含量估計為系統輸出,該當前 素點的準確含量估計反饋到該卡爾曼濾波器,在該卡爾 濾波器中,用作為下一個像素點含量估計的新初始值,該 卡爾曼濾波器的輸出用作為該遺傳算法分解模塊的起始點 以便加速遺傳算法的進化; (心6)移到下一個像素點,該卡爾曼濾波器接收來 =特徵聽賴塊的正交雜,來自獻氣補償模塊的 當前像素獅jf:值,以及來自該遺傳算法分賴塊的前一 個像素朗含f料,並姻甸曼驗麟對該當前像 素點進行頻譜分解,來自該遺傳算法分解模塊的準確含量 前像 曼 該 --------- (請先閎讀背面之注意事項再填寫本頁) 線 -26- 冰張u適用中國固家2敝297公瘦).Scope of patent application 5 15 Printed by the Intellectual Property Bureau of the Ministry of Economic Affairs, Consumer Cooperatives, 20 Classification data of neural networks, and the estimation of the first-pixel content, and the content of the first pixel. Technology pixels: ^ :: spectrum decomposition. The waver turns out the visit of the current orthogonal contact scale processing module, and obtains the measurement value of the current pixel from the big louse repairer module, the data from the network, and the Kalman filter. Content estimation, and use the genetic algorithm to decompose the current spectrum accurately to determine the content estimation. The Kalman II = itB is used as a compilation algorithm to resolve the shortcomings of the algorithm in order to accelerate the content estimation.遽 Wave 2 is used as the new initial value of the next __ pixel points; 下) Move to the next __ pixel point, the card receiver receives the difference of the j-characteristics module from the compensation module of the domain. Tianyue i pixel point measurement 1 value comes from the classification data of the artificial neural network, and from the decomposition method of the traditional method to estimate the content, and the current pixel is spectrally decomposed by using the Carr fork filter technology. The accurate content estimation of the genetic abandonment decomposition module is used in the Kalman filter _ used as the previous pixel _ frequency meter to determine the content vector of the current point; (d-7) Return to step (μ) , Repeat the steps (d-5) and (d-6) up to the last pixel in the hyperspectral characteristic cube. 10 · As described in item 9 of the scope of the patent application, "Analysis of the hyperspectral map and pixels --------- Xiang (please read the notes on the back before filling in this # 1. Page)} -2 'This paper uses China National Standards (CNS) (210x297 mm) A8 B8 C8 D8 patent application frequency ° knife solution method ", where in step (d-3), the first pixel is calculated The parameter estimator for the content estimation of points is a least squares (LS) estimator. 11. "Hyperspectral image analysis and pixel frequency solution method" as described in the patent application_item 9, where in Turn to ⑷), and the parameter estimator for estimating the leaf in s of the first pixel is a maximum likelihood (g) estimator. 12: According to the "Hyperspectral Image Analysis and Pixel Frequency Decomposition Method 4" described in item 9 of the scope of the patent application, its order is in step ⑷) +, and the parameter estimator for calculating the content estimation of the first pixel is- Evolutionary algorithm. 13f filed a patent_ item "High" Image Analysis and Pixel Point Frequency Resolution Method ", where step (d) is composed of the following steps: 15 Printed by the Consumers' Cooperative of the Intellectual Property Bureau of the Ministry of Economic Affairs 20, ( d 1) receiving the spectral characteristics of the material of interest from the spectral library, and sub-characteristics, the frequency features from the crane spectral library are orthogonally normalized in the feature preprocessing module, Thus, the feature space is decomposed into a set of orthogonal feature sets; (d 2) the first pixel point content estimation module receives the orthogonal features from the feature preprocessing chess block 'and a first-level first pixel from the atmospheric compensation module Point measurement data, and use a parameter estimator to calculate the content estimate of the first pixel point, except for the first pixel point, the first pixel point contains the material, and the compensation is calculated from the money compensation model. Some measurement data; (d 3) the Karl Cry / Wave filter successively obtained these positive signs from the feature pre-processing module, from the pixel measurement in the domain before complementing the pixels. CNS) A4 (21Qx297 J / J patent application range 15 A8 B8 C8 D8 20 data printed by the Consumer Cooperatives of the Central Standards Bureau of the Ministry of Economic Affairs, and the content estimation of the content vector estimation module of the first pixel, and the current pixel frequency spectrum using Kalman filtering technology Decomposition, the content estimate of the first pixel is used as the initial value of the Kalman filter, and the Kalman filter is used to output the content estimate of the current pixel to the decomposing algorithm decomposition module; &quot; &quot; (d-4) Evaluate the estimation error of the content estimation value of the current pixel, &amp; when the estimation error is greater than a preset value, go to step (Heart 5), where the telephone error is given by the covariance matrix of the Kalman filter, otherwise, Go to step (d-6), the genetic anomaly decomposition module only bypasses the content estimator obtained by the Kalman filter; (d-5) the genetic anomaly decomposition module receives the orthogonal features of the feature pre-processing module , The current pixel measurement value from the atmospheric compensation module, and the current pixel content estimation from the Kalman filter, and using the genetic algorithm to the current image The points are subjected to accurate spectral decomposition to obtain the f-quasi-peak content estimate, which is the system output. The accurate content estimate of the current prime point is fed back to the Kalman filter, which is used as the next pixel point in the Kalman filter. The new initial value of the content estimate. The output of the Kalman filter is used as the starting point of the decomposition module of the genetic algorithm to accelerate the evolution of the genetic algorithm. (Heart 6) Move to the next pixel and the Kalman filter receives = Feature listens to the orthogonal noise of the block, the current pixel jf: value from the donation compensation module, and the previous pixel from the genetic algorithm's block contains the f data, and the current pixel Point for spectral decomposition, the exact content from the genetic algorithm decomposition module is like Manga --------- (Please read the precautions on the back before filling this page) Line-26- Bing Zhang u for China Gu family 2 敝 297 male thin). 估計在該卡®曼濾波器中用作為前_個.像素關含量估 計,以便準確估計該當前點的含量矢量;以及 (d-7)回到步驟(d-4),在該高譜特性立方體内重復 遞推步驟(d-4),(d-5)與(d-6)直到最後一個像素點者。 5 U·如申請專利範圍第13項所述之「高讓圖象分析及像素點 頻譜分解方法」,其中在步驟⑷)中,計算該第一像 素點的含量估計的參數估計器是一個最小二乘(LS)估計 器者。 15.如申請專利範圍第13項所述之「高譜圖象分析及像素點 10頻譜分解方法」,其中在步驟(d-2)中,計算該第一像 素點的含量估計的參數估計器是一個極大似然(ML)估 計器者。 16·如申請專利範圍第13項所述之「高譜圖象分析及像素點 頻譜分解方法」,其中在步驟(d-2)中,計算該第一像 15素點的含量估計的參數估計器是一個進化算法者。 17.如申請專利範圍第1項所述之「高譜圖象分析及像素點 頻谱分解方法」,其中步驟(d)由以下步驟構成: 經濟部中央標隼局員工消費合作社印製 (請先閎讀背面之注意事項再填寫本頁) 、1T (心1)接枚來自頻譜庫的感興趣材料的頻譜特徵, 並在特徵預處理模塊中預處理這些特徵,來自該頻譜庫的 20這些頻譜特徵在該特徵預處理模塊中進行正交歸一化,從 而將該特徵空間分解成一组正交特徵集合; (心2)人工神經網絡接收來自該特徵預處理模塊的 這些正交特徵以及來自該大氣補償模魂的一组第一像素點 測量數據,並進行材料分類選擇,該人工神經網絡繞過從 -27- 本紙張纽適财關家制1 ( eNS } Α4_ ( 2ι()χ2970 &gt; 經濟部中央標隼局員工消費合作社印製 503375 A8 B8 C8 ____08____ 六、申請相範目 —— — 該大氣補償模塊到後續模塊的這些測量數據; jd-3)第一像素點含量估計模塊接收來自該特徵預 處理模2的雜正交特徵,來自獻氣補健塊的第一像 素點測ΐ數據,以及來自該人工神經網絡的第一像素點的 5 ,類,據,並利用-個參數估計器計算該第-像素點的含 1矢量,該第一像素點的含量估計用作為卡爾曼濾波器的 ,始值,對於除了該第一像素點以外的像素點,該第一像 素點含S估計模塊僅僅繞過來自該人工神經網絡模塊的這 些測量數據; 10 (°^4)該卡爾曼濾波器接收來自該特徵預處理模塊 的k些正交特徵,來自該大氣補償模塊的當前像素點測量 數據,來自該人工神經網絡的分類數據,以及該苐一像素 點的^量矢量估計模塊的含量估計,並用卡爾曼濾波技術 進行當前像素點的頻譜分解,該卡爾曼濾波器輸出該當前 15像素點的含量估計到遺傳算法分解模塊; 一(心5)評價該當前像素點的含量估計值的估計誤差, 當該估計誤差大於預設值時,轉到步驟(d-6),其中估 計誤差由卡爾曼濾波器的協方差矩陣給出,否則,轉到步 驟(水7),該遣傳算法分解模塊僅僅繞過由卡爾曼濾波器 20 何·到的含量估計; (d&lt;)遺傳算法分解模塊接收該特徵預處理模塊的 正交特徵’來自該大氣補償模塊的該當前像素點測量值, 來自該人工神經網絡的分類數據,以及來自該卡爾曼濾波 器的當前像素點的含量估計,並利用該遺傳算法對該當前 -28- -----------Φ! (請先閲讀背面之注意事項再填寫本頁) 、1TThe estimation is used as the first_pixel-off content estimate in the Kalman filter in order to accurately estimate the content vector of the current point; and (d-7) returns to step (d-4) in the hyperspectral characteristic Repeat the steps (d-4), (d-5) and (d-6) up to the last pixel in the cube. 5 U · As described in the "Method for High-resolution Image Analysis and Pixel Spectrum Decomposition" described in item 13 of the scope of the patent application, in step ii), the parameter estimator for calculating the content estimation of the first pixel is a minimum The squarer (LS) estimator. 15. The "hyperspectral image analysis and pixel 10 spectrum decomposition method" as described in item 13 of the scope of patent application, wherein in step (d-2), a parameter estimator for estimating the content of the first pixel Is a maximum likelihood (ML) estimator. 16. The "Hyperspectral Image Analysis and Pixel Spectrum Decomposition Method" as described in item 13 of the scope of patent application, wherein in step (d-2), the parameter estimation of the content estimation of the 15 prime points of the first image is calculated Is an evolutionary algorithmist. 17. The "Hyperspectral Image Analysis and Pixel Spectrum Decomposition Method" as described in item 1 of the scope of patent application, wherein step (d) consists of the following steps: Printed by the Consumer Cooperatives of the Central Bureau of Standards of the Ministry of Economic Affairs (please First read the notes on the back and then fill out this page), 1T (Heart 1) one after another the spectral features of the material of interest from the spectral library, and pre-process these features in the feature pre-processing module, 20 of these from the spectral library The spectral features are orthogonally normalized in the feature pre-processing module to decompose the feature space into a set of orthogonal feature sets; (heart 2) the artificial neural network receives these orthogonal features from the feature pre-processing module and A set of first pixel measurement data of the atmospheric compensation model soul, and material classification selection, the artificial neural network bypasses from -27- this paper New Zealand financial system 1 (eNS} Α4_ (2ι () χ2970 &gt; Printed by the Consumer Cooperatives of the Central Bureau of Standards of the Ministry of Economic Affairs 503375 A8 B8 C8 ____08____ VI. Application Phases — These measurement data from the atmospheric compensation module to subsequent modules; jd-3 ) The first pixel content estimation module receives the hetero-orthogonal feature from the feature pre-processing module 2, the first pixel measurement data from the Qi donation block, and the 5 from the first pixel of the artificial neural network. , Class, data, and use a parameter estimator to calculate the 1-pixel-containing vector, the content estimate of the first pixel is used as the Kalman filter, the initial value, except for the first pixel Pixels, the first pixel containing the S estimation module only bypasses these measurement data from the artificial neural network module; 10 (° ^ 4) the Kalman filter receives k orthogonalities from the feature preprocessing module Characteristics, the current pixel measurement data from the atmospheric compensation module, the classification data from the artificial neural network, and the content estimation of the first pixel vector estimation module, and the spectrum of the current pixel using Kalman filtering technology Decomposed, the Kalman filter outputs the content estimate of the current 15 pixels to the genetic algorithm decomposition module; one (heart 5) evaluates the content estimate of the current pixel When the estimated error is greater than a preset value, go to step (d-6), where the estimated error is given by the covariance matrix of the Kalman filter, otherwise, go to step (water 7), the report The algorithm decomposition module only bypasses the content estimation by the Kalman filter 20; (d &lt;) The genetic algorithm decomposition module receives the orthogonal features of the feature pre-processing module 'the current pixel measurement from the atmospheric compensation module Value, the classification data from the artificial neural network, and the current pixel content estimate from the Kalman filter, and use the genetic algorithm to the current -28- ----------- Φ! (Please read the notes on the back before filling this page) 、 1T 申明專利範圍 10 15 經濟部中央標率局員工消費合作社印製 20 像素點進行準確頻譜分解以得出準確含量估計,該卡爾曼 濾波器輸出用作為該遺傳算法分解模塊的起始點以便加速 遣傳算法的進化,該準確含量估計反饋到該卡爾曼濾波 器’在該卡爾曼濾波器中,用作為下一個像素點含量估計 的新初始值; 乂d-7)移到下一個像素點,該卡爾曼濾波器接收來 自該特徵預處理模塊的正交特徵,來自該大氣補償模塊的 當前像素點測量值,來自該人工神經網絡的分類數據,以 及來自該遺傳算法分解模塊的前一個像素點的含量估計, 並利用卡爾曼濾波技術對該當前像素點進行頻譜分解,來 自該遺傳算法分解模塊的準確含量估計在該卡爾曼濾波器 中用作為前一個像素點的含量估計,以便準確估計該當前 點的含量矢量;以及 (d-8)回到步驟(d_5),在該高譜特性立方體内重復 遞推步驟(心5)與(d-6)直到最後一個像素點者。 18= 如申請專利範圍第17項所述之「高譜圖象分析及像素點 頻譜分解方法」’其中在步驟⑷)中,計算該第-像 素點的含量估計的參數估計器是一個最小二乘(LS)估計 器者。 19·如申請專利範圍第17項所*之「高譜分析及像素點 頻語分解方法」,其中在步驟⑷)+,計算該第一像 素點的含量估計的參數估計器是-個極大似然(ML)估 計器者。 20·如申清專利範圍第17項所述之「高譜圖象分析及像素點 —裝·1 (請先閲讀背面之注意事項再填寫本頁) 、11 線 本紙張U適用中國國家標 -29- 503375 ABCD 經濟部中夬標準局員工消費合作社印製 六、申請專利範圍 頻譜分解方法」,其中在步騍(d-3)中,計算該第一像 素點的含量估計的參數估計器是一個進化算法者。 21·如申請專利範圍第3項所述之「高譜圖象分析及像素點 頻譜分解方法」,其中該卡爾曼濾波器是一個萬能魯棒卡 5爾曼漶波器者。 22·如申請專利範圍第21項所述之「高譜圖象分析及像素點 頻譜分解方法」,其中該萬能魯棒濾波器由以下步驟組成: (f.l)接收來自該大氣補償模塊的該像素點的該測量 數據; 10 (f·2)在模糊邏輯模塊中,利用模糊邏輯推理方法檢 驗該測量數據,在此,該模糊邏輯推理方法依據豐富的模 糊邏輯規則有選擇地決定拒絕該測量數據或校正該測量數 據,或承認該測量數據; (f.3)輸出該校正後的測量數據或錯誤標誌到預處理 15 模塊,在此,該預處理模塊執行狀態轉移矩陣和測量矩障 的計算; (f.4)將該狀態轉移矩陣從該預處理模塊送到狀態矢 量預測模塊,將前一個狀態矢量從狀態矢量更新模塊送到 該狀態矢量預測模塊,在此,該狀態矢量預測模塊進行狀 20 態矢量預測,即該下一個像素點的含量估計; (f.5)將該狀態轉移矩陣從該預處理模塊送到協方差 傳播模塊,在此,該協方差傳播模塊計算當前估計誤差的 協方差; (16)將該測量矩唪及當前測量矢量從該預處理模塊 -30- 本紙張;顧中關家鮮(CNS ) ( 210X297公羡) -- ----------Φ--------1T------^φί (請先閲讀背面之注意事項再填寫本頁) 503375 8 8 88 ABCD 六 經濟部中央標準局員工消費合作社印製 申請專利範圍 送到測量殘差計算模塊,在此,該測量殘差計算模塊接收 來自該狀態矢量預測模塊的該狀態矢量預測值,並通過從 該當前測量矢量中減去該測量矩陣與該狀態矢量預測值的 乘積來計算測量殘差; 5 (f.7)將該當前估計誤差的協方差從該協方差傳播模 塊送到最優增益計算模塊,在此,該最優增益計算模塊計 算最優增益; (f.8)將該最優增益從該最優增益計算模塊送到協方 差更新模塊,在此,該協方差更新模塊更新該估計誤差的 10 協方差; (f.9)將該估计誤差的協方差更新值從該協方差更新 模塊送到協方差傳播模塊;以及 (〇〇)將該最優增益從該最優增益計算模塊送到該 狀態矢量更新模塊,在此,該狀態矢量更新模塊接收來自 15 該測量殘差計算模塊的該測量殘差,並進行狀態矢量更 新,也即,該下一個像素點的含量估計者。 23·如申請專利範圍第22項所述之「高譜圖象分析及像素點 頻譜分解方法」,其中步驟(f.2)由以下步驟組成: (f.2-1)將該測量數據送到一個模糊器模塊,在此, 20該模糊器執行標量映射,也就是把該測量數據的範圍轉換 到一個相應的論域,並進行模糊化,也就是把該測量數據 變換成適當的語義值,這些語義值被標為模糊集合,並對 所得的模糊輸入用模糊集合及其隸屬函數([〇,!])解釋確 定的測量數據; -31- 本纸張尺度適用中國國家標準(CNS ) A4規格(210X297公釐) —-------------、玎------ (請先閱讀背面之注意事項再填寫本頁} 503375 經濟部中央標準局員工消費合作社印製 A8 B8 C8 D8 六、申請專利範圍 (f-2-2)將該模糊輸入從該板糊器模境送到一個模糊 推理機構,在此,該权糊推理機構本質上模仿人類的決策 機理,並利用模糊推理規則推出模糊輸出,來自模糊規則 庫的該权糊邏輯推理規則借助一組語義規則歸納了專家的 5目標和策略,該模糊規則庫由應用域知識及目標組成;以 及· (£2-3)將該模糊輸出從該模糊推理機構送到一個模 糊還原模塊,在此,模糊還原模境生成一個破定的有效測 量數據,該數據最好地表達了推理模糊輸出的可能分布 10者。 24.如申請專利範圍第5項所述之「高譜圖象分析及像素點 頻4分解方法」,其中該遺傳算法分解由以下步驟組成: (1) 一個編竭模塊隨機生成而一,组二進制字符串集 合,這組二進制字符串代表與該高譜圖象立方體的該像素 15點相關的該含量估計,遺傳算法對這組二進制字符亊進行 操作,且該組二進制字符_被送到一個解碼模塊; (2) 該解碼模塊對這組二進制字符串進行解碼,該 解媽模塊的輸出即是關於該超級頻譜圖象立方體像素點的 該含量估計集合,這組含量估計集合被送到一値適合度計 20异模塊,該含量估計給出了該像素點中所包含的每個感興 趣材料的百分比; (3) 適合度計算模塊計算每個含量估計的適合度值, 在該適合度計算模塊中,指標函數取該二進制字符串(也 稱之為染色體),並返回一個值,然後把該指標函數的值 -32- 本纸張尺度逋用中國國家標準(CNS ) A4規格(210X297公釐) ----------Φ,------1T------M.#— (請先閎讀背面之注意事項再填寫本頁) A8 B8 C8 -----------— D8申請專利範圍 經濟部中央標準局員工消費合作社印製 適應該遣傳算法,該適合度值是基於由該 字ϋ代表的所有可能解的性能的回答 ’該編碼予符串的 Ml#該適合度值也越高,該適合度值再送到 一個繁殖模塊; 5 匕(4)執行判別過程以確定是否終止進化,在此,判 別才曰標足義為總進化代數,當該遣傳算法遞推到該總進化 代數時’ _其#-個具有最大適合度值的二進制字符串 作為解並且該遣傳算法退出進化,其相應的含量估計矢量 就疋該像素點的含量估計; 10 (5)在該繁殖模塊中,基於來自該適合度計算模塊 的該適合麟tB,進膽i,纽,鮮歧基於適者生 存的規律,即,這些適合度值高的二進制字符串會在新一 代中有大量複製,一旦這些二進制字符串被繁殖或複製作 為下一代使用,該二進制字符_將在配對集合中進行另外 15兩類操作,即交配與變異,從而繁殖; (6) 在交配模塊中,通過交換該二進制字符串(染 色體)的頭和尾來形成子字符串集合,交配為該二進制字 符串提供一種機制通過隨機過程來混合和匹配其期望的品 質,在該隨機過程中,首先,從該繁瘦模塊形成的該匹配 20集合中選出兩個新生成的字符串;其二,沿這兩個字符串 一致地隨機選擇一個交換位置;其三,交換該交換位置以 後的所有字符; (7) 在變異模塊中,偶而改變在一個特定字符串位 置上的值,該變異是對任何簡單位的恆久損失的一種保險 本纸張尺度適用中國國家標準(CNS ) A4規格(210X 297公釐) (請先閲讀背面之注意事項再填寫本頁) .裝&lt; T 線 申請專利範圍 15 經濟部中央標準局員工消費合作社印製 20 措施,該變異的發生概率極低, 只有一個字符串發生變異;以及 =制&quot;麵㈣合制轉瑪模塊’ 5tt^()’(4)’(5)’(6),(7),以及⑻步驟執行者。 LUt·概邮24項賴之「高顧_彳及像素點 °曰刀法J ’其中該遺傳算法分解由以下步驟組成: _ (1) 一個編碼模塊隨機生成一組二進制字符串集合, =级二進^1封㈣代麵鶴觸象立方體的該像素點相 P的該3里估计,遺傳算法對這組二進制字符串進行操 作,且該組二進制字符串被送到一個解碼模塊; 、(2)該解碼模塊對這組二進制字符串進行解碼,該 解竭模塊的如即是關於該超級頻譜縣立謂像素點的 ,含讀t憎合,賴含量估計集合被❹卜個適合度計 算模塊,該含量估計給出了該像素點中所包含的每個感興 趣材料的百分比者; (3) 適合度計算模塊計算每個含量估計的適合度值, 在該適合度計算模塊t,指標函數取該二進制字符事(也 稱之為染色體),並返回一個值,然後把該指標函數的值 峡射為適合度以適應該遺傳算法,該適合度值是基於由該 子符串代表的所有可能解的性能的回答,該編碼字符串的 該含量估計越好,該適合度值也越高,該適合度值再送到 一個繁殖模塊者; (4) 執行判別過程以確定是否終止進化,通過評價 該一進制字符串間的差別來進行判別,當該字符串間的差 以至在字符串集合中平均 1·1 (請先閱讀背面之注意事項再填寫本頁) 訂 線 -34- 經濟部中夬標隼局〃貝工消費合作社印製 503375 A8 B8 C8 一丨丨 ..................-............... -… ~__ D8 六、申請專利範圍— '^ : 別小於-個預設值時,該遺傳算法退出進化,在此之後, 選,一個適合度值最大的二進制字符串作為解,其相應的 合量估計矢量就是該像素點的含量估計; (5)在該繁殖模塊中,基於來自該適合度計算模塊 5的該適合度輪出,進行繁殖,在此,該龍是基於適者生 存的規律,即,這些適合度值高的二進制字符串會在新一 代中有大量複製,一旦這些二進制字符串被繁殖或複製作 為下一代使用,該二進制在配對集合中進行另外兩類操 作,即交配與變異,從而繁殖; 10 (6)在交配模塊中,通過交換該二進制字符串(染 色體)的頭和尾來形成子字符串集合,交配為該二進制字 符串提供一種機制通過隨機過程來混合和匹配其期望的品 質,在該隨機過程t,首先,從該繁殖模塊形成的該匹配 集合中選出兩個新生成的字符串;其二,沿這兩個字符$ 15 一致地隨機選擇一個交換位置;其三,交換該交換位置以 後的所有字符; (?)在變異模塊令,偶而改變在一個特定字符串位 置上的值,該變異是對任何簡單位的恆久損失的一種保險 措施,該變異的發生概率極低,以至在字符串集合令平均 20只有一個字符串發生變異;以及 (8)將該新的二進制字符串集合送到該解碼模塊, 然後按(2),(3),(4),(5),(6),(7),以及⑻步驟執行者。 26·如申請專利範圍第25項所述之「高譜圖象分析及像素點 頻譜分解方法」,其中步驟(f.2)由以下步驟組成: -35- 本纸張尺度適用中國國家標準(CNS ) A4規格(210X297公釐) (請先閱讀背面之注意事項再填寫本頁} 、11 線 —____________ 08503375 A8 B8 C8 六、申請專利範圍 經濟部中央標準局員工消費合作社印製 (£2-1)將該測量數據送到一個模糊器模塊,在此, 該模糊器執行標量映射,也就是_測量數據的範圍轉換 到-個相應的論域,並進行模糊化,也就是她測量數據 變換成適當的浯義值。這些語義值被標為模糊集合,並對 5所得的模掬輸入用模糊集合及其隸屬函數([oj])解 定的測量數據; ’ (f.2-2)將該模糊輸入從該模糊器模塊送到一個模糊 推理機構,在此,該模糊推理機構本質上模仿人類的決策 機理,並利用模糊推理規則推出模糊輸出,來自模糊規= 10庫的該模糊邏輯推理規則借助一組語義規則歸納了專家的 目標和策略,該模糊規則庫由應用域知識及目標組成;以 及 (f.2-3)將該模糊輸出從該模糊推理機構送到一個模 糊還原模塊,在此,模鞠還原模塊生成一個確定的有效測 15量數據,該數據最好地表達了推理模糊輸出的可能分布 者。 27·如申請專利範圍第9項所述之「高譜圖t分析及像素點 頻譜分解方法」,其中該卡爾曼濾波器是一萬能魯棒濟波 器者。 〜 20 28·如申請專利範圍第27項所述之「高譜圖象分析及像素點 頻4刀解方法」,其申該萬能魯棒濾波器由以下步課組成: (Π)接收來自該大氣補償模塊的該像素點的該測量 數據; (£2)在模糊邏輯模塊中,利用模輞邏輯推理方法檢 I. , — (請先閎讀背面之注意事項再填寫本I) 、-0 線 -36- 本紙張尺度適用中國國家榡準(CNS ) A4規格(210X297公釐) 503375 A8 B8 C8 D8 申請專利範圍 15 經濟部中央標準局員工消費合作社印製 驗該測量減,在此,該_邏娜财法域豐富的模 糊邏輯規财選獅蚊拒賴測藏校正該測量數 據,或承認該測量數據; (〇)輸出該校正後的測量數據或錯誤標誌到預處理 模塊,在此,該預處理模塊執行狀態轉移矩陣和測量矩陣 的計算; (£4)將該狀態轉移矩陣從該預處理模塊送到狀態矢 量預測模塊,將前一個狀態矢量從狀態矢量更新模塊^到 該狀態矢量預測模塊,在此,該狀態矢量預測模塊進行狀 態矢量預測,即該下一個像素點的含量估計; (f.5)將該狀態轉移矩陣從該預處理模塊送到協方差 傳播模塊,在此,該協方差傳播模塊計算當前估計誤差的 協方差; (f.6)將該測量矩陣及當刖測量矢量從該預處理模塊 送到測量殘差計算模塊,在此,該測量殘差計算模塊接收 來自該狀態矢量預測模塊的該狀態矢量預測值,並通過從 該當前測量矢量中減去該測量矩陣與該狀態矢量預測值的 乘積來計算測量殘差; (f.7)將該當前估計誤差的協方差從該協方差傳播模 塊送到最優增益計算模塊,在此,該最優增益計算模塊計 算最優增益; (£8)將該最優增益從該最优增益計算模塊送到協方 差更新模塊,在此,該協方差更新模塊更新該估計誤差的 協方差; -37- 本纸張尺度適用中國國家橾準(CNS ) A4規格(210 X 297公釐) (請先閱讀背面之注意事項再填寫本頁) -裝 訂 線Declared patent scope 10 15 Printed 20 pixels by the Consumer Cooperative of the Central Standards Bureau of the Ministry of Economic Affairs for accurate spectral decomposition to obtain accurate content estimates. The output of the Kalman filter is used as the starting point of the genetic algorithm decomposition module to speed up the process. The evolution of the algorithm, the accurate content estimate is fed back to the Kalman filter ', in the Kalman filter, it is used as the new initial value for the content estimate of the next pixel; 乂 d-7) move to the next pixel, The Kalman filter receives orthogonal features from the feature pre-processing module, current pixel measurement values from the atmospheric compensation module, classification data from the artificial neural network, and previous pixel points from the genetic algorithm decomposition module The content estimation of the current pixel is performed using Kalman filtering technology, and the current pixel is spectrally decomposed. The accurate content estimation from the genetic algorithm decomposition module is used as the content estimation of the previous pixel in the Kalman filter in order to accurately estimate the Content vector at the current point; and (d-8) returns to step (d_5), where the high-spectrum characteristic stands Repeat in vivo recursive step (heart 5) and (d-6) until the last pixel person. 18 = As described in the "Method of Hyperspectral Image Analysis and Pixel Spectrum Decomposition" described in item 17 of the scope of the patent application, where in step ii), the parameter estimator for calculating the content estimation of the -pixel point is a least square Multiplier (LS) estimator. 19. As in the "spectrum analysis and pixel frequency decomposition method" of item 17 of the scope of patent application, in step ⑷) +, the parameter estimator for calculating the content estimation of the first pixel is a very similar (ML) Estimator. 20 · As stated in item 17 of the scope of the patent application, “Hyperspectral Image Analysis and Pixels—Packing · 1 (Please read the precautions on the back before filling out this page), 11 thread paper U applies Chinese national standard- 29- 503375 ABCD printed by the Consumers' Cooperative of the China Standards Bureau of the Ministry of Economic Affairs. 6. The spectrum decomposition method of patent application scope. "In step (d-3), the parameter estimator for calculating the content estimate of the first pixel is An evolutionary algorithmist. 21. The "Hyperspectral Image Analysis and Pixel Spectrum Decomposition Method" as described in item 3 of the scope of the patent application, wherein the Kalman filter is a universal and robust Kalman waver. 22. The "Hyperspectral Image Analysis and Pixel Spectrum Decomposition Method" as described in item 21 of the scope of patent application, wherein the universal robust filter consists of the following steps: (fl) receiving the pixel from the atmospheric compensation module 10 (f · 2) In the fuzzy logic module, use fuzzy logic inference to check the measurement data. Here, the fuzzy logic inference method selectively decides to reject the measurement data based on the rich fuzzy logic rules. Or correct the measurement data, or acknowledge the measurement data; (f.3) output the corrected measurement data or error flags to the pre-processing module 15, where the pre-processing module performs the calculation of the state transition matrix and the measurement moment barrier (F.4) The state transition matrix is sent from the preprocessing module to the state vector prediction module, and the previous state vector is sent from the state vector update module to the state vector prediction module. Here, the state vector prediction module performs State 20 state vector prediction, that is, the content estimation of the next pixel; (f.5) The state transition matrix is sent from the preprocessing module to the covariance Broadcast module, here, the covariance propagation module calculates the covariance of the current estimated error; (16) the measurement moment and the current measurement vector from the preprocessing module -30- this paper; Gu Zhongguan Jiaxian (CNS) (210X297 public envy)----------- Φ -------- 1T ------ ^ φί (Please read the precautions on the back before filling this page) 503375 8 8 88 ABCD Six consumer economic cooperatives of the Central Standards Bureau of the Ministry of Economic Affairs printed the patent application scope and sent it to the measurement residual calculation module. Here, the measurement residual calculation module receives the state vector prediction value from the state vector prediction module and passes Subtract the product of the measurement matrix and the predicted value of the state vector from the current measurement vector to calculate the measurement residual; 5 (f.7) send the covariance of the current estimation error from the covariance propagation module to the optimal gain A calculation module, where the optimal gain calculation module calculates the optimal gain; (f.8) sending the optimal gain from the optimal gain calculation module to the covariance update module, where the covariance update module updates 10 covariance of the estimated error; (f.9) The covariance update value of the estimated error is sent from the covariance update module to the covariance propagation module; and (〇〇) the optimal gain is sent from the optimal gain calculation module to the state vector update module, where the state The vector update module receives the measurement residual from the measurement residual calculation module and performs a state vector update, that is, the content estimator of the next pixel. 23. The "Hyperspectral Image Analysis and Pixel Spectrum Decomposition Method" as described in item 22 of the scope of the patent application, wherein step (f.2) consists of the following steps: (f.2-1) send the measurement data To a fuzzer module, here, the 20 fuzzer performs scalar mapping, that is, the range of the measurement data is converted to a corresponding universe, and fuzzification is performed, that is, the measurement data is transformed into an appropriate semantic value These semantic values are labeled as fuzzy sets, and the fuzzy inputs and their membership functions ([0 ,!]) are used to interpret the determined measurement data for the fuzzy inputs obtained; -31- This paper scale applies Chinese National Standards (CNS) A4 specifications (210X297 mm) —-------------, 玎 ------ (Please read the precautions on the back before filling out this page} 503375 Staff Consumption of Central Standards Bureau, Ministry of Economic Affairs Cooperative prints A8 B8 C8 D8 6. The scope of patent application (f-2-2) sends the fuzzy input from the stenter model to a fuzzy inference mechanism. Here, the right inference mechanism mimics the human Decision mechanism and use fuzzy reasoning rules to derive fuzzy input The fuzzy logic reasoning rule from the fuzzy rule base summarizes the 5 goals and strategies of the expert by means of a set of semantic rules. The fuzzy rule base consists of application domain knowledge and goals; and (£ 2-3) the fuzzy output From the fuzzy reasoning mechanism to a fuzzy reduction module, the fuzzy reduction model generates a valid and broken measurement data, which best expresses the possible distribution of the inferred fuzzy output of 10. 24. If the scope of patent application The "Hyperspectral Image Analysis and Pixel Frequency 4 Decomposition Method" described in item 5, wherein the genetic algorithm decomposition consists of the following steps: (1) A compilation module is randomly generated and a set of binary strings, which The set of binary strings represents the content estimate related to the 15 points of the pixel of the hyperspectral image cube, the genetic algorithm operates on the set of binary characters 亊, and the set of binary characters _ is sent to a decoding module; (2) The decoding module decodes the set of binary strings, and the output of the solution module is about the content of the pixels of the super-spectrum image cube. This set of content estimation set is sent to a 20-meter fitness module. The content estimation gives the percentage of each material of interest contained in the pixel; (3) The fitness calculation module calculates each The fitness value of the content estimate. In the fitness calculation module, the indicator function takes the binary string (also called chromosome) and returns a value, and then the value of the indicator function is -32- this paper scale逋 Use Chinese National Standard (CNS) A4 specification (210X297 mm) ---------- Φ, ------ 1T ------ M. # — (Please read the back first (Please note this page before filling out this page) A8 B8 C8 ------------- D8 Patent Application Scope The Central Consumers Bureau of the Ministry of Economic Affairs's Consumer Cooperatives printed and adapted the repatriation algorithm. The fitness value is based on The answer to the performance of all possible solutions represented by the word 'The Ml # of the coded string is the higher the fitness value, and the fitness value is sent to a breeding module; 5 (4) Perform a discrimination process to determine whether Stop evolution. Here, it is judged that the standard meaning is the total evolution algebra. When the algorithm is recursive, When the total evolution algebra '_ ## is a binary string with the largest fitness value as the solution and the retransmission algorithm exits evolution, its corresponding content estimation vector will be the content estimation of the pixel; 10 (5) in In the breeding module, based on the fitness tB from the fitness calculation module, enter the i, new, and fresh, based on the law of survival of the fittest, that is, there will be a large number of binary strings with high fitness values in the new generation. Copy, once these binary strings are reproduced or copied for use as the next generation, the binary character _ will perform another 15 two types of operations in the pairing set, namely, mating and mutation, thereby multiplying; (6) In the mating module, by swapping The head and tail of the binary string (chromosome) form a set of substrings, and mating provides a mechanism for the binary string to mix and match its desired qualities through a random process. In this random process, first, from the traditional Two newly generated strings are selected from the set of matching 20 formed by the thin module; secondly, one is randomly selected along the two strings consistently Third, exchange all characters after the exchange position; (7) In the mutation module, occasionally change the value at a specific string position, the mutation is a insurance paper against the permanent loss of any simple bit Zhang scale is applicable to China National Standard (CNS) A4 specification (210X 297 mm) (Please read the precautions on the back before filling out this page). Applicable scope of patent for T-line 15 Printed by the Consumers' Cooperative of the Central Standards Bureau of the Ministry of Economic Affairs 20 measures, the probability of occurrence of the mutation is extremely low, and only one character string is mutated; and the system of “combination system” 5 tt ^ () '(4)' (5) '(6), (7 ), And who performed the step. LUt · Email 24 items of "Guo Gu_ 彳 and pixel °° knife method J 'where the genetic algorithm decomposition consists of the following steps: _ (1) An encoding module randomly generates a set of binary string sets, = level The ^ 1 estimation of the binary phase of the pixel-phase phase P of the face-to-face crane crane cube is 3, and the genetic algorithm operates on the set of binary strings, and the set of binary strings is sent to a decoding module;, ( 2) The decoding module decodes this set of binary strings, and the depletion module is about the super-spectrum county pixels, including reading t, and the content estimation set is calculated based on the fitness. Module, the content estimate gives the percentage of each material of interest contained in the pixel; (3) the fitness calculation module calculates the fitness value of each content estimate, and in the fitness calculation module t, the index The function takes the binary character thing (also called chromosome), and returns a value, and then casts the value of the indicator function into a fitness to adapt to the genetic algorithm. The fitness value is based on the sub-symbol string. All The answer to the performance of the possible solution, the better the content of the encoded string is estimated, the higher the fitness value is, and the fitness value is sent to a breeding module person; (4) the judgment process is performed to determine whether to terminate the evolution. Evaluate the difference between the unsigned strings to make a judgment. When the difference between the strings is an average of 1 · 1 in the set of strings (please read the precautions on the back before filling this page). Printed by the Bureau of Standards and Industry, Cooperated with Shellfish Consumer Cooperative 503375 A8 B8 C8 丨 丨 .................... .. -... ~ __ D8 VI. Patent Application Scope-'^: When it is less than-a preset value, the genetic algorithm exits evolution. After that, choose a binary string with the highest fitness value as the solution. The corresponding total amount estimation vector is the content estimate of the pixel; (5) In the breeding module, based on the fitness from the fitness calculation module 5, the rotation is performed and breeding is performed. Here, the dragon is based on the survival of the fittest The rule is that these binary strings with high fitness values will There are a large number of copies. Once these binary strings are reproduced or copied for use as the next generation, the binary performs two other types of operations in the paired set, namely mating and mutation, thereby multiplying; 10 (6) In the mating module, by exchanging the The head and tail of a binary string (chromosome) form a set of substrings. Mating provides a mechanism for the binary string to mix and match its desired qualities through a random process. At this random process t, first, from the breeding module Two newly generated strings are selected from the formed matching set; secondly, a swap position is consistently and randomly selected along the two characters $ 15; third, all characters after the swap position are swapped; (?) In mutation The module orders, occasionally changing the value at a specific string position. This mutation is a insurance measure against the permanent loss of any simple bit. The probability of this mutation is extremely low, so that there is only one string in the set of strings. Mutated; and (8) send the new set of binary strings to the decoding module, and then press (2) (3), (4), (5), (6), (7), and by ⑻ steps. 26. The "Hyperspectral Image Analysis and Pixel Spectrum Decomposition Method" as described in item 25 of the scope of the patent application, wherein step (f.2) consists of the following steps: -35- This paper scale applies Chinese national standards ( CNS) A4 specification (210X297 mm) (Please read the precautions on the back before filling out this page} 、 11 line —____________ 08503375 A8 B8 C8 6. Application for patents Printed by the Consumer Cooperative of the Central Standards Bureau of the Ministry of Economic Affairs (£ 2- 1) The measurement data is sent to a fuzzer module, where the fuzzer performs scalar mapping, that is, the range of _ measurement data is converted to a corresponding domain, and fuzzification is performed, that is, her measurement data transformation Into appropriate semantic values. These semantic values are labeled as fuzzy sets, and the measured data obtained from the fuzzy set and its membership function ([oj]) are used to input the model data obtained from 5; '(f.2-2) The fuzzy input is sent from the fuzzer module to a fuzzy inference mechanism. Here, the fuzzy inference mechanism essentially imitates human decision-making mechanism, and uses fuzzy inference rules to derive fuzzy output from the fuzzy rule = 10 The fuzzy logic reasoning rules of the library summarize the experts' goals and strategies with the help of a set of semantic rules. The fuzzy rule base consists of application domain knowledge and goals; and (f.2-3) sends the fuzzy output from the fuzzy reasoning mechanism. To a fuzzy reduction module, here, the Juju reduction module generates a determined effective measurement of 15 data, which best represents the possible distribution of the inferred fuzzy output. 27. As described in item 9 of the scope of patent application "Hyperspectral t analysis and pixel spectral decomposition method", in which the Kalman filter is a universal robust wave filter. ~ 20 28 · "Hyperspectral image analysis" as described in item 27 of the scope of patent application And pixel point frequency 4-knife solution method ", which claims that the universal robust filter consists of the following steps: (Π) receiving the measurement data of the pixel point from the atmospheric compensation module; (£ 2) in the fuzzy logic module In the test, I use the logic reasoning method of the mold rim to check I., — (Please read the precautions on the back before filling in this I), -0 line -36- This paper size applies to China National Standards (CNS) A4 specifications (210X297) %) 503375 A8 B8 C8 D8 Patent application scope 15 The Consumer Cooperative of the Central Standards Bureau of the Ministry of Economic Affairs has printed and tested the measurement. Here, the ___________________________________ Or acknowledge the measurement data; (〇) output the corrected measurement data or error flag to the preprocessing module, where the preprocessing module performs the calculation of the state transition matrix and the measurement matrix; (£ 4) the state transition matrix The preprocessing module is sent to the state vector prediction module, and the previous state vector is updated from the state vector update module to the state vector prediction module. Here, the state vector prediction module performs state vector prediction, that is, the next pixel point. Content estimation; (f.5) sending the state transition matrix from the preprocessing module to the covariance propagation module, where the covariance propagation module calculates the covariance of the current estimation error; (f.6) the measurement matrix And when the measurement vector is sent from the preprocessing module to the measurement residual calculation module, the measurement residual calculation module receives the prediction vector from the state vector. The state vector predicted value, and calculate the measurement residual by subtracting the product of the measurement matrix and the state vector predicted value from the current measurement vector; (f.7) the covariance of the current estimation error from the The variance propagation module is sent to the optimal gain calculation module, where the optimal gain calculation module calculates the optimal gain; (£ 8) The optimal gain is sent from the optimal gain calculation module to the covariance update module, here , The covariance update module updates the covariance of the estimated error; -37- This paper size applies to China National Standard (CNS) A4 (210 X 297 mm) (Please read the precautions on the back before filling this page )-Gutter 經濟部中央標準局員工消费合作社印製 ^ (f·9)將該估計誤差的協方差更新值從該協方差更新 模塊送到協方差傳播模塊;以及 將該最優增益從該最優增益計算模塊送到該 狀恶矢里更新模塊,在此,該狀態矢量更新模塊接收來自 5該測里殘差计异模塊的該測量殘差,並進行狀離失量更 新,也即,該下-個像素點的含量估計者。^ 29·如申請專繼圍第28項所述之「高觸象分析及像素點 頻譜分解方法」,其中步驟(f.2)由以下步驟組成: (f.2-1)將該測量數據送到一個模糊器模塊,在此, 10該模糊器執行標量映射,也就是把該測量數據的範圍韓換 到一個相應的論域,並進行模糊化,也就是把該測量數據 變換成適畲的語義值,這些語義值被標為模糊集合,並對 所得的模糊輸入用模糊集合尽其隸屬函數([^])解釋確 定的測量數據; 15 (f·2—2)將該模翔輸入從該模糊器模塊送到一個模糊 推理機構,在此,該模糊推理機構本質上模仿人類的決策 機理’並利用模糊推理規則推出模糊輸出,來自模糊規則 庫的該模糊邏輯推理規則借助一組語義規則歸納了專家的 目才示和取略’該模糊規則庫由應兩域知識及目標组成;以 20 及 (f.2-3 )將該模糊輸出從該模糊推理機構送到一個模 糊還原模塊,在此,模糊還原模塊生成一個確定的有效測 里數據,該數據袁好地表達了推理模构輸出的可能分布 者0 -38- 本纸張尺度適用中國國家樣隼(CNS ) A4現格(210X297公釐) --------------II------ (請先閎讀背面之注意事項再填寫本頁;&gt; 經濟部智慧財產局員工消费合作社印製 Α8 Β8 __ C8 ^—---^ D8申請專利範圍 ~—---— 30·如申請專利範圍第13項所述之「高譜圖象分析及像素點 頻譜分解方法」,其中該卡爾曼濾波器是一萬能魯棒滹波 器者。 31·如申請專利範圍第30項所述之「高譜圖象分析及像素點 5頻譜分解方法」,其中該萬能魯棒濾波器由以下步驟組成: (f.l)接收來自該大氣補償模塊的該像素點的該測量 數據; (f.2)在模糊邏輯模塊中,利用模糊邏輯推理方法檢 驗該測量數據,在此,該模糊邏輯推理方法依據豐富的模 10糊邏輯規則有選擇地決定拒絕該測量數據或校正該測量數 據,或承認該測量數據; (f」)輸出該校正後的測量數據或錯誤標諸到預處理 模塊,在此,該預處理模塊執行狀態轉移矩陣和測量矩陣 的計算; 15 (f·4)將該狀態轉移矩陣從該預處理模塊送到狀態矢 量預測模塊’將前一個狀態矢量從狀態矢量更新模塊送到 該狀態矢量預測模塊’在此,該狀態矢量預測模塊進行狀 態矢量預測,即該下一個像素點的含量估計; (f.5)將該狀態轉移矩陣從該預處理模塊送到協方差 20 傳播模塊,在此,該協方差傳播模塊計算當前估計誤差的 協方差, (£6)將該測量矩陣及當前測量矢量從該預處理模塊 送到測量殘差計算模塊,在此,該測量殘差計算模塊接收 來自該狀態矢量預測模塊的該狀態矢量預測值,並通過從 •39- 本纸張又度適用中國國家梂準(CNS ) Α4規格(210Χ:297公釐) (請先閎讀背面之注意事項再填寫本頁) -裝j 、1T 線 刈3375 A8 B8 C8 D8 申請專利範圍 10 15 經濟部智慧財產局員工消費合作社印製 20 該當前測量矢量中減去該測量矩陣·與該狀態矢量預測值的 乘積來計算測量殘差; (〇)將該當前估計誤差的協方差從該協方差傳播模 塊送到最優增益計算模塊,在此,該最優增益計算模塊計 算最優增益; (f.8)將該最優增益從該最優增益計算模塊送到協方 差更新模塊,在此,該協方差更新模塊更新該估計誤差的 協方差; (f.9)將該估計誤差的協方差更新值從該協方差更新 模塊送到協方差傳播模塊;以及 (Π0)將該最優增益從該最優增益計算模塊送到該 狀態矢量更新模塊,在此,該狀態矢量更新模塊接收來自 該測量殘差計算模塊的該測量殘差,並進行狀態矢量更 新,也即,該下一個像素點的含量估計者。 32.如申請專利範圍第31項所述之「高譜圖象分析及像素點 頻譜分解方法」,其中步驟(f.2)由以下步驟組成: (f.2-1)將該測量數據送到一個模糊器模塊,在此, 該模糊器執行標量映射,也就是把該測量數據的範圍轉換 到一個相應的論域,並進行模鞠化,也就是柄該測量數據 變換成適當的語義值,這些語義值被標為模糊集合,並對 所得的模糊輸入用模糊集合及其隸屬函數([Ο』)解釋確 定的測量數據; 5 (£2-2)將該模糊輸入從該模糊器模塊送到一個模糊 推理機構,在此,該模翔推理機構本質上模仿人類的決策 秦 (請先閲讀背面之注意事項再填寫本頁) -裝I 訂 良紙張尺度適用中國國家榡準(CNS) 公釐y 5 5 ;以 機=:=推理規則推出_輪出,來自模糊規則 庫的該t糊邏她理規則借助—組語義規购納了專家的 目標和策略’該模_規則庫由顧域知識及目標組成 及 $、、、 (£2-3)將該獅輸出從賴糊推理機構送到一個模 糊還原概,在此,模_原模塊生成—辦定的有效測 量數據,雜據最瓶表達了推理_輪$的可能分布 者0 33·如申請專利範圍第η項所述之r高譜圖t分析及像素點 10 頻“77解方/i:」’其t該卡目曼濾波II是—萬能魯棒滅波 器者。 ^ 34.如申請專利範圍苐33項所述之「高譜圖I分析及像素點 頻譜分解方法」,其中該萬能魯棒濾波器由以下步驟組成: (f.l )接收來自該大氣補償模塊的該像素點的該測量 15 數據; (f.2)在模糊邏輯模塊令,利用模糊邏輯推理方法檢 驗該測量數據;在此,該模糊邏輯推理方法依據豐富的模 糊邏輯規則有選擇地決定拒絕該測量數據或校正該測量數 據’或承認該測量數據; 2〇 (G)輸出該校正後的測量數據或錯誤標誌到預處理 模塊,在此,該預處理模塊執行狀態轉移矩陣和測量矩障 的計算; U4)將該狀態轉移矩陣從該預處理模塊送到狀態矢 量預測模塊,將前一個狀態矢量從狀態矢量更新模塊送到 -41- 本纸法Χϋλ用中國國家梂準(⑽)入视格(加乂297公着) 503375 經濟部智慧財產局員工消費合作社印制衣 A8 B8 C8 D8 六、申請專利範圍 該狀態矢量預測模塊,在此,該狀態矢量預測模塊進行狀 態矢量預測,即該下一個像素點的含量估計者; (£5)將該狀態轉移矩陣從該預處理模塊送到協方差 傳播模塊,在此,該協方差傳播模塊計算當前估計誤差的 5協方差; (f.6)將該測量矩陣及當前測量矢量從該預處理模塊 送到測量殘差計算模塊,在此,該測量殘差計算模塊接收 來自該狀態矢量預測模塊的該狀態矢量預測值,並通過從 該當前測量矢量中減去該測量矩陣與該狀態矢量預測值的 10 乘積來計算測量殘差; (f.7)將該當前估計誤差的協方差從該協方差傳播模 塊送到最優增益計算模塊,在此,該最優增益計算模塊計 算最優增益; (f.8)將該最優增益從該最優增益計算模塊送到協方 15差更新模塊,在此,該協方差更新模塊更新該估計誤差的 協方差; (f.9)將該估計誤差的協方差更新值從該協方差更新 模塊送到協方差傳播模塊;以及 (f.10)將該最優增盈從該最優增益計算模塊送到該 20 狀態矢量更新模塊,在此,該狀態矢量更新模塊接收來自 該測量殘差計算模塊的該測量殘差,並進行狀態矢量更 新’也即,該下一個像素點的含量估計者。 35·如申請專利範圍第34項所述之「高譜圖象分析及像素點 頻譜分解方法」,其中步驟(f.2)由以下步驟組成: -42- 本紙張尺度適用中國國家標準(CNS)A4規格(210 X 297公釐) -----------裝------—訂 i (請先閱讀背面之注意事項再填寫本頁) # k 州375Printed by the Consumer Cooperative of the Central Standards Bureau of the Ministry of Economic Affairs ^ (f · 9) The covariance update value of the estimated error is sent from the covariance update module to the covariance propagation module; and the optimal gain is calculated from the optimal gain The module is sent to the state vector update module. Here, the state vector update module receives the measurement residual from the 5 residual measurement difference module, and updates the state deviation amount, that is, the next- Pixel estimator. ^ 29. As described in the application for the "High Touch Image Analysis and Pixel Spectrum Decomposition Method" described in item 28, step (f.2) consists of the following steps: (f.2-1) the measurement data Send it to a fuzzer module, where 10 the fuzzer performs scalar mapping, that is, the scope of the measurement data is changed to a corresponding domain, and fuzzification is performed, that is, the measurement data is transformed into a suitable These semantic values are labeled as fuzzy sets, and the fuzzy inputs are used to explain the determined measurement data with their fuzzy membership functions ([^]) on the fuzzy inputs; 15 (f · 2-2) input the model From this fuzzer module to a fuzzy inference mechanism, where the fuzzy inference mechanism essentially imitates human decision-making mechanisms and uses fuzzy inference rules to derive fuzzy outputs. The fuzzy logic inference rules from the fuzzy rule base rely on a set of semantics The rule summarizes the expert's eyesight and omissions. The fuzzy rule base consists of two domains of knowledge and goals; the fuzzy output is sent from the fuzzy inference mechanism to a fuzzy reduction model with 20 and (f.2-3). Here, the fuzzy reduction module generates a certain valid survey data. This data Yuan Hao expresses the possible distribution of the output of the inference model. 0 -38- This paper scale is applicable to China National Sample (CNS) A4. (210X297 mm) -------------- II ------ (Please read the notes on the back before filling out this page; &gt; Consumer Consumption Cooperative of Intellectual Property Bureau, Ministry of Economic Affairs Printed A8 Β8 __ C8 ^ —--- ^ D8 patent application scope ~ --- --- 30. "Hyperspectral image analysis and pixel spectral decomposition method" as described in item 13 of the patent application scope, where The Kalman filter is a universal holstering wave filter. 31. The "hyperspectral image analysis and pixel 5 spectral decomposition method" as described in item 30 of the scope of patent application, wherein the universal robust filter consists of the following The steps are composed of: (fl) receiving the measurement data of the pixel point from the atmospheric compensation module; (f.2) in the fuzzy logic module, using fuzzy logic inference method to check the measurement data, here, the fuzzy logic reasoning method Selectively reject the measurement based on the abundance of logic rules Data or correct the measurement data, or acknowledge the measurement data; (f ") output the corrected measurement data or errors to the preprocessing module, where the preprocessing module performs the calculation of the state transition matrix and the measurement matrix; 15 (f · 4) The state transition matrix is sent from the preprocessing module to the state vector prediction module 'send the previous state vector from the state vector update module to the state vector prediction module' Here, the state vector prediction module performs State vector prediction, that is, the content estimation of the next pixel; (f.5) The state transition matrix is sent from the preprocessing module to the covariance 20 propagation module, where the covariance propagation module calculates the current estimation error. Covariance, (£ 6) The measurement matrix and the current measurement vector are sent from the preprocessing module to the measurement residual calculation module, where the measurement residual calculation module receives the state vector prediction value from the state vector prediction module , And passed from • 39- this paper is again applicable to China National Standards (CNS) Α4 specifications (210 ×: 297 mm) (Please read the note on the back first (Please fill in this page again)-Install j, 1T line 3375 A8 B8 C8 D8 Patent application scope 10 15 Printed by the Intellectual Property Bureau of the Ministry of Economic Affairs Employee Cooperatives 20 Print out the current measurement vector minus the measurement matrix and predict the state vector The product of the values is used to calculate the measurement residual; (0) The covariance of the current estimated error is sent from the covariance propagation module to the optimal gain calculation module, where the optimal gain calculation module calculates the optimal gain; (f .8) send the optimal gain from the optimal gain calculation module to the covariance update module, where the covariance update module updates the covariance of the estimated error; (f.9) the covariance of the estimated error The updated value is sent from the covariance update module to the covariance propagation module; and (Π0) sends the optimal gain from the optimal gain calculation module to the state vector update module, where the state vector update module receives from the The measurement residual of the measurement residual calculation module performs state vector update, that is, the content estimator of the next pixel. 32. The "Hyperspectral Image Analysis and Pixel Spectrum Decomposition Method" as described in item 31 of the scope of patent application, wherein step (f.2) consists of the following steps: (f.2-1) send the measurement data To a fuzzer module, where the fuzzer performs scalar mapping, that is, transforms the range of the measurement data into a corresponding universe of speech, and performs modularization, that is, the measurement data is transformed into appropriate semantic values These semantic values are labeled as fuzzy sets, and the fuzzy inputs and their membership functions ([0]) are used to interpret the determined measurement data for the obtained fuzzy inputs; 5 (£ 2-2) remove the fuzzy inputs from the fuzzer module Send to a fuzzy reasoning mechanism, here, this model Xiang reasoning mechanism essentially imitates human decision-making Qin (please read the precautions on the back before filling this page)-installed I The well-defined paper size applies to China National Standards (CNS) Y 5 5 mm; Introduced by machine =: = inference rules_roll out, the t logic rules from the fuzzy rule base acquired expert goals and strategies with the help of group semantic rules. Gu domain knowledge and target composition And $ ,,, (£ 2-3) sent the lion output from the Laibu reasoning mechanism to a fuzzy reduction scheme. Here, the module _ original module generated-the effective measurement data, and the reasoning expressed the reasoning. The possible distribution of round $ 0 33. As described in the r scope of the patent application, the r hyperspectral map t analysis and pixel 10-frequency "77 solution / i:" 'its t this Kahman filter II is- Universal Robust Wavebreaker. ^ 34. The "Hyperspectral I Analysis and Pixel Spectrum Decomposition Method" as described in the scope of application for patent 33 items, wherein the universal robust filter consists of the following steps: (fl) receiving the atmospheric compensation module from the 15 measurements of pixels; (f.2) In the fuzzy logic module, the fuzzy logic inference method is used to check the measurement data; here, the fuzzy logic inference method selectively decides to reject the measurement based on rich fuzzy logic rules. Data or correct the measurement data 'or acknowledge the measurement data; 20 (G) outputs the corrected measurement data or error flag to the pre-processing module, where the pre-processing module performs the calculation of the state transition matrix and the measurement moment barrier U4) The state transition matrix is sent from the pre-processing module to the state vector prediction module, and the previous state vector is sent from the state vector update module to the -41- paper method. (Publication 297) 503375 Printed clothing A8 B8 C8 D8 by the Consumer Cooperatives of the Intellectual Property Bureau of the Ministry of Economy Module, here, the state vector prediction module performs state vector prediction, that is, the content estimator of the next pixel; (£ 5) sends the state transition matrix from the preprocessing module to the covariance propagation module, here, The covariance propagation module calculates 5 covariances of the current estimation error; (f.6) sends the measurement matrix and the current measurement vector from the preprocessing module to the measurement residual calculation module, where the measurement residual calculation module receives The state vector prediction value from the state vector prediction module, and calculate a measurement residual by subtracting a 10 product of the measurement matrix and the state vector prediction value from the current measurement vector; (f.7) the current estimation The covariance of the error is sent from the covariance propagation module to the optimal gain calculation module, where the optimal gain calculation module calculates the optimal gain; (f.8) sends the optimal gain from the optimal gain calculation module Go to the covariance update module, where the covariance update module updates the covariance of the estimated error; (f.9) update the covariance of the estimated error from the covariance update module To the covariance propagation module; and (f.10) sending the optimal gain from the optimal gain calculation module to the 20 state vector update module, where the state vector update module receives the measurement residual calculation module The measurement residual and update the state vector ', that is, the content estimator of the next pixel. 35. The "Hyperspectral Image Analysis and Pixel Spectrum Decomposition Method" as described in item 34 of the scope of the patent application, wherein step (f.2) consists of the following steps: -42- This paper scale is applicable to the Chinese National Standard (CNS ) A4 size (210 X 297 mm) ----------- install ------- order i (Please read the precautions on the back before filling this page) # k 州 375 申請專利範圍 10 15 經濟部智慧財產局員工消費合作社印製 20 (£2-1)將該測量數據送到一個模糊器模塊,在此, 該模糊ϋ猜標量_,樣是把_量輯的範圍轉換 到-個相應的論域,並進浦糊化,也就是把铜量數據 變換成適當的#義值,這些語義值被標為模糊集合,並對 所得的模糊輸人用翻集合及其隸屬聽(⑽)解 定的測量數據; (f.2-2)將該模糊輸入從該模糊器模塊送到一個模糊 推理機構,在此,該模糊推理機構本質上模仿人類的決策 •機理,並利用模糊推理規則推出模糊輸出,來自模糊規則 庫的該模糊邏輯推理規·助—錄義規崎納了專家的 目標和策略,該模蝴規則庫由應用域知識及目標組成;以 及 &gt; (f.2o)將該模糊輸出從該模糊推理機構送到 糊還原模塊,在此,模糊還原模塊生成一個確定的有效測 量數據,該數據最好地表$了推理模糊輸出的可能分布 者。 36.如申請專利範圍第3項所述之「高細象分析及像素點 頻譜分解方法」,其令該遺傳算法分解由以下步驟組成: (1) 一個編碼模塊隨機生成而一組二進制字符串集 合,這組二進制字符串代表與該高譜圖象立方體的該像素 點相關的該含量估計,遺傳算法對這組二進制字符串進行 操作,且該組一進制子符串被送到一個解碼模塊; (2) 該解碼模塊對這組二進制字符串進行解碼, 解碼模塊的輸出即是關於該超級頻譜圖象立方體像素點 該 的 (請先閎讀背面之注意事項再填寫本頁} -裝‘ 訂 線 -43- 503375 經濟部智慧財產局員工消费合作社印製 A8 B8 C8 ................................................—&quot;&quot;&quot;—&quot;&quot;&quot; -..... .. ................ , ____i 13S 六、申請專利範圍~ —-— =含量估計集合,這組含量估計集合被送到一個適合度計 算模塊,該含量估計給出了該像素點令所包含的每個感興 趣材料的百分比,· 八 (3) 適3度计异拉塊計异每個含量估計的適合度值, 5在該適合度計算模塊令,指標函數取該二進制字符串(也 稱之為染色體),錢回—健,織把翻標函數的值 $射為適合度以適麟前算法,該適合度值是基於由該 字符串代表的所有可能解的性能的回答,該編碼字符串的 該含量估計越好,該適合度值也越高,該適合度值再送到 10 —個繁瘦模塊; (4) 執行判別過程以確定是否終止進化,在此,判 別指標定義為總進化代數,當該遺傳算法遞推到該總進化 代數時,選擇其令一個具有最大適合度值的二進制字符串 作為解並且該遺傳算法退出進化,其相應的含量估計矢量 15 就是該像素點的含量估計; (5) 在該繁殖模塊中,基於來自該適合度計算模塊 的該適合度輸出,進行繁殖,在此,該繁殖是基於適者生 存的規律,即,這些適合度值高的二進制字符串會在新一 代中有大量複製,一旦這些二進制字符串被繁殖或複製作 20 為下一代使用,該二進制字符串將在配對集合中進行另外 兩類操作,即交配與變異,從而繁殖; (6) 在交配模塊中,通過交換該二進制字符串(染 色體)的頭和尾來形成子字符串集合,交配為該二進制^ 符串提供一種機制通過隨機過程來混合和匹配其期望的品 άφ (請先聞讀背面之注意事項再填寫本頁) -........I------ g I II 1........ · 1—1 —I— I ml 、1T 線 -44- 本紙張尺度適用中國國家標準(CNS ) Α4規格(210Χ297公庚)Scope of patent application 10 15 Printed by the Consumer Cooperatives of the Intellectual Property Bureau of the Ministry of Economic Affairs 20 (£ 2-1) This measurement data is sent to a fuzzer module. Here, the fuzzer guesses the scalar _. The range is converted to a corresponding universe of discourse, and is further gelatinized, that is, the copper data is transformed into appropriate #semantic values. These semantic values are labeled as fuzzy sets. (F.2-2) Send the fuzzy input from the fuzzer module to a fuzzy inference mechanism, where the fuzzy inference mechanism essentially imitates human decision-making mechanism, And use fuzzy inference rules to derive fuzzy output. The fuzzy logic inference rules from the fuzzy rule base assist the dictation rules to capture the goals and strategies of experts. The model rule base is composed of application domain knowledge and goals; and &gt; ( f.2o) The fuzzy output is sent from the fuzzy inference mechanism to the paste reduction module. Here, the fuzzy reduction module generates a certain valid measurement data, which best represents the possible analysis of the inferred fuzzy output. Cloth. 36. The "high-resolution analysis and pixel spectral decomposition method" described in item 3 of the scope of the patent application, which causes the genetic algorithm decomposition to consist of the following steps: (1) a coding module randomly generates a set of binary strings Set, the set of binary strings represents the content estimate related to the pixel point of the hyperspectral image cube, the genetic algorithm operates on the set of binary strings, and the set of unary substrings is sent to a decoder Module; (2) The decoding module decodes the set of binary strings, and the output of the decoding module is about the pixels of the super spectrum image cube (please read the precautions on the back before filling this page)- '' Order-43- 503375 Printed by A8 B8 C8 ... ........— &quot; &quot; &quot;-&quot; &quot; &quot; -..... .. .......... ......, ____i 13S Sixth, the scope of patent application ~ ----= content estimation set, this set of content estimation set is sent to a fitness calculation module, the content estimation gives the The percentage of each material of interest contained in the pixel order, eight (3) suitable for 3 degrees, different for the pull block, different for each content estimated fitness value, 5 in the fitness calculation module order, the index function is taken as Binary string (also known as chromosome), Qian Huijian, weaving the value of the rescaling function $ to fit the previous algorithm, the fitness value is based on all possible solutions represented by the string Performance answer. The better the content of the encoded string is, the higher the fitness value is. The fitness value is then sent to 10 thin modules. (4) The judgment process is performed to determine whether to terminate the evolution. The discriminant index is defined as the total evolution algebra. When the genetic algorithm is recursive to the total evolution algebra, it is selected to make a binary string with the largest fitness value as the solution and the genetic algorithm exits evolution, and its corresponding content estimation vector 15 is the content estimation of the pixel; (5) in the breeding module, breeding is performed based on the fitness output from the fitness calculating module, and here, the breeding is based on the survival of the fittest The rule is that these binary strings with high fitness values will be copied in large numbers in the new generation. Once these binary strings are reproduced or copied as 20 for the next generation, the binary string will be used in the pairing set for another two. (6) In the mating module, the head and tail of the binary string (chromosome) are exchanged to form a substring set, and mating provides a mechanism for the binary ^ symbol string. Random process to mix and match their desired products (please read the notes on the back before filling out this page) -........ I ------ g I II 1 ..... ... · 1-1 —I— I ml, 1T line -44- This paper size is applicable to China National Standard (CNS) Α4 specification (210 × 297 cm) 申請專利範圍 10 15 經濟部智慧財產局員工消費合作社印製 20 質’在該隨機過程中,首參,彡 集:中選出兩個新生成的字符串?其二 擇-個交換位置;其三,交換該交換= 在變異模塊巾,偶·變在—雜定字符串位 ς场值。異是對任何簡單位的蚊損失的_種保險 口= ’該變異的發生概率極低,以至在字料集合中平均 &quot;有一個字符串發生變異,以及 1⑻將該新的二進制字符串集合送到該解媽模塊, 後按(2) ’(3) ’(4),(5),⑹’⑺,以及⑻步驟執行者。 4申清專利範圍第5項所述之「高谱圖象分析及像素點 頻譜分解方法」,其中該遺鮮法分解由以下步驟組成: (1) 一個編碼模塊隨機生成而一組二進制字符串集 合,碰二進制字符串代表與該高譜圖象立方體的該像素 點相關的1¾含量估計,遺彳轉法對這纟H制字符串進行 操作,且該組一進制字符串被送到一個解瑪模逸; (2) 該解碼模塊對這組二進制字符_進行解碼,該 解碼模塊的輸出即是關於該超級頻譜圖象立方體像素點的 該含量估計集合,這組含量估計集合被送到一偭適合度計 算模塊,該含量估計給出了該像素點中所包含的每個感興 趣材料的百分比; (3) 適合度計算模塊計算每個含量估計的適合度值, 在該適合度計算模塊中,指標函數取該二進制字符串(也 稱之為染色體),並返回一個值,然後把該指標函數的值 -45- 本纸張又度逋用中國國家棣準(CNS ) A4規格(210X297公着) 請 先 閲 讀 背 之 注 意 事 項 再 填 窝 本&lt; 頁&lt; ,射為適合度以適應該遺傳算法,該適合度值是基於由該 子符串代表的所有可能解的性能的回答,該編碼字符串的 該含量估計越好,該適合度值也越高,該適合度值再送到 一個繁殖模堍; (4) 執行判別過程以嫁定是否終止進化,在此,判 別指標定義為總進化代數,當該遺傳算法遞推到該總進化 代數時,選擇其t一個具有最大適合度值的二進制字符串 作為解並且該遺傳鼻法退出進化,其相應的含量估計矢量 就是該像素點的含量估計; (5) 在該繁殖模塊中,基於來自該適合度計算模塊 的該適合錄出,断倾,在此,該紐絲於適者生 存的規律,即,這些適合度值高的二進制字符_會在新一 代中有大S複製,一旦這些二進制字符串被繁殖或複製作 為下一代使用,該二進制字符串將在配對集合令進行另外 兩類操作,即交配與變異.,從而繁殖; (6) 在父配模塊中,通過交換該二進制字符串(染 色體)的頭和尾來形成子字符串集合,交配為該二進财 符串提供一種機制通過隨機過程來混合和匹配其期望的品 質,在該隨機過程中,首先,從該繁賴塊形成的該匹配 集合尹選出兩简生錢字符串;其二,沿這兩個字符串 -致地隨機選擇-個交換位置;其三,交換該交換位置以 後的所有字符; ⑺在變異模塊中,躺改變在·—轉定字符串位 置上的值,該變異是對任何料位_久損失的—種保險 5 A8 B8 C8 D8 申請專利範圍 措施,該變異的發生概率極低,以至在字符串集合中平均 只有一個字符串發生變異;以及 (8)將該新的二進制字符串集合送到該解碼模塊, 然後按(2) ’(3),(4),(5),⑹,⑺,以及(8)步驟執行者。 5 38·如申請專利範圍第9項所述之「高譜圖象分析及像素點 頻譜分解方法」,其令該遣傳算法分解由以下步驟組成·· (1) 一個編碼模塊隨機生成而一組二進制字符串集 合,這組二進制字符串代表與該高譜圖象立方體的該像素 點相關的該含量估計,遺傳算法對這組二進制字符串進行 ίΟ操作,且該組一進制字符串被送到一個解碼模塊; (2) 該解碼模塊對這組二進制字符串進行解碼,該 解碼模塊的輸出即是關於該超級頻譜圖象立方體像素點的 該含量估計集合,這組含量估計集合被送到一個適合度計 算模塊,該含量估計給出了該像素點中所包含的每個感興 15 趣材料的百分比; ' (3) 適合度計算模塊計算每個含量估計的適合度值, 在該適合度計算模塊中,指標函數取該二進制字符串(也 稱之為染色體),並返回一個值,然後把該指標函數的值 經濟部中央標隼局員工消費合作社印製 20 =射為適合度輯應錢·法,該適合舰是基於由該 字代表的所有可能解的性能的回答,該編碼字符串的 該各里估计越好,該適合度值也越高,該適合度值再送到 一個繁殖模塊; (4) 執行判別過程以確定是否終止進化,在此,判 別指標定義為總進化代數,當該遺傳算法遞推到該總進化 503375 A8 B8 C8 D8 六、申請專利範圍 經濟部智慧財產局員工消費合作社印製 =,選擇其中-個具有最大適合度值的二進制字符串 作,解並域遺傳算法如進化,其相應的含量估計矢量 就是該像素點的含量估計; mtr繁雜塊中,基牡自該適合度計算模塊 5 ^該適j度輸',進行繁瘦’在此’該繁殖是基於適者生 子的規/ gp ’廷些適合度值高的二進制字符串會在新一 代中有大量複製,-旦這些二進游符串㈣M複製作 為下代使用該—進制字符串將在配對集合中進行另外 兩類操作,即交配與變異,從而繁殖; 10 (6)在交配模+ ’通過交換該二進纟彳字符串(染 色體)的頭和尾來形成子字符串集合,交配為該二進制字 符串提供-種機制通過隨機過程來混合和匹配其期望的品 質,在該隨機過財,首先,㈣繁賴塊職的該匹配 集合中選出兩個新生成的字符串;其二,沿這兩個字符串 15 -致崎機選擇-個交換位置;其三,交換鼓換位置以 後的所有字符; (7)在變異模塊中,偶而改變在一個特定字符串位 置上的值,該變異是對任何簡單位的恆久損失的一種保險 拾施,战變異的發生概率極低,以至在字符串集合中平均 20只有一個字符串發生變異;以及 、(8)將該新的一進制字符串集合送到該解碼模塊, 然後按(2) ’(3),(4),(5),(6),(7),以及⑻步驟執行者。 39.如申請專利範圍第13項所紅「高譜_分析及像素點 頻譜分解方法」,其中該遺傳算法分解由以下步驟組成: -48- 本纸張尺度適用中國國家榡準(CNS ) A4規格(210X297公釐) (請先閎讀背面之注意事項再填寫本頁} -裝 訂 線Scope of patent application 10 15 Printed by the Intellectual Property Bureau of the Ministry of Economic Affairs and Consumer Cooperatives 20 qualitative 'In this random process, the first reference, the collection: choose two newly generated strings? The second choice-an exchange position; the third , Exchange this exchange = in the mutation module towel, even · change in-heterodyne string bit field value. The difference is that any kind of simple bit of mosquitoes is safe. = 'The probability of occurrence of this mutation is extremely low, so that there is an average mutation in the word set, and 1⑻ the new binary string set Send it to the solution module, and then press (2) '(3)' (4), (5), ⑹'⑺, and ⑻ step performer. 4 Apply for the "Hyperspectral Image Analysis and Pixel Spectrum Decomposition Method" described in Item 5 of the Patent Scope, where the legacy method decomposition consists of the following steps: (1) A coding module randomly generates a set of binary strings Set, the binary string represents the 1¾ content estimate related to the pixel point of the hyperspectral image cube, the widow's transformation method operates on the H string, and the set of unary strings is sent to a Xie Ma Mo Yi; (2) The decoding module decodes the set of binary characters _, and the output of the decoding module is the content estimation set about the pixel points of the cube of the super-spectrum image. Fitness calculation module, the content estimate gives the percentage of each material of interest contained in the pixel; (3) the fitness calculation module calculates the fitness value of each content estimate, in the fitness calculation module The indicator function takes the binary string (also called the chromosome) and returns a value, and then the value of the indicator function is -45. This paper again uses the Chinese national standard (CNS) A4 specification (210X297) Please read the precautions before filling in the book <Page>. The fitness is adapted to the genetic algorithm. The fitness value is based on the sub-string. The answer to the performance of all possible solutions, the better the content of the encoded string is estimated, the higher the fitness value is, and the fitness value is sent to a breeding model; (4) the discriminating process is performed to marry whether to terminate the evolution Here, the discriminant index is defined as the total evolution algebra. When the genetic algorithm recurses to the total evolution algebra, it selects a binary string with the maximum fitness value as the solution and the genetic nose method exits evolution, and its corresponding The content estimation vector of is the content estimation of the pixel; (5) In the breeding module, based on the fitness recording from the fitness calculation module, the dip is dumped. Here, the button is based on the law of survival of the fittest, that is, These binary characters with high fitness values will have a large S copy in the new generation. Once these binary strings are reproduced or copied for the next generation, the binary word Strings will undergo two other types of operations in the mating set order, namely mating and mutation. Thus, breeding; (6) In the parental matching module, the substring set is formed by swapping the head and tail of the binary string (chromosome), Mating provides a mechanism for the binary financial string to mix and match its desired qualities through a random process. In the random process, first, two simple money-generating strings are selected from the matching set Yin formed by the promiscuous block; Second, randomly select a swap position along these two strings; third, swap all characters after the swap position; ⑺ In the mutation module, lie and change the value at the position of the transposed string , This mutation is for any material level _ long loss-a kind of insurance 5 A8 B8 C8 D8 patent scope measures, the probability of this mutation is extremely low, so that only one string in the set of strings will mutate on average; and (8 ) Send the new set of binary strings to the decoding module, and then perform steps (2) '(3), (4), (5), ⑹, ⑺, and (8). 5 38. According to the "Hyperspectral Image Analysis and Pixel Spectrum Decomposition Method" described in item 9 of the scope of the patent application, the decomposition algorithm decomposition consists of the following steps ... (1) An encoding module is randomly generated and A set of binary strings, the set of binary strings representing the content estimate related to the pixel point of the hyperspectral image cube, the genetic algorithm performs a 0 operation on the set of binary strings, and the set of unary strings is Sent to a decoding module; (2) the decoding module decodes the set of binary strings, and the output of the decoding module is the content estimation set about the pixels of the super-spectrum image cube, and the content estimation set is sent To a fitness calculation module, the content estimate gives the percentage of each interesting material contained in the pixel; '(3) The fitness calculation module calculates the fitness value of each content estimate. In the degree calculation module, the indicator function takes the binary string (also called chromosome) and returns a value, and then the value of the indicator function is evaluated by Printed by the Ministry of Standards and Technology Bureau's Consumer Cooperatives 20 = Shooting as a suitability appropriation method, which is a response based on the performance of all possible solutions represented by the word. Well, the fitness value is also higher, and the fitness value is sent to a breeding module again. (4) The discrimination process is performed to determine whether to terminate the evolution. Here, the discrimination index is defined as the total number of evolutionary generations. The total evolution 503375 A8 B8 C8 D8 VI. Patent application scope Printed by the Consumer Cooperatives of the Intellectual Property Bureau of the Ministry of Economic Affairs =, select one of them as a binary string with the largest fitness value, and solve the genetic algorithm such as evolution, its corresponding The content estimation vector of the pixel is the content estimation of the pixel; in the mtr complex block, the basic fitness calculation module 5 ^ the appropriate j degree is lost, and the thinning is performed. Here, the reproduction is based on the rules of the fittest child. gp 'These binary strings with high fitness values will have a large number of copies in the new generation,-once these binary escape strings ㈣ M copy as the next generation to use this-the base string will Two other types of operations are performed in the paired set, that is, mating and mutation, so as to reproduce; 10 (6) In the mating mode + 'by forming the substring set by swapping the head and tail of the binary string (chromosome), mating Provide a mechanism for the binary string to mix and match its desired qualities through a random process. First, in the random wealth, two new generated strings are selected from the matching set of the block; second, , Along these two strings 15-Zhiqi machine selection-a swap position; Third, swap all characters after the drum change position; (7) In the mutation module, occasionally change the value at a specific string position, This mutation is a kind of insurance against the permanent loss of any simple bit. The probability of occurrence of war mutation is extremely low, so that only one string in the set of strings will mutate on average; and (8) the new one Send the set of string of characters to the decoding module, and then press (2) '(3), (4), (5), (6), (7), and the step performer. 39. For example, "High-spectrum_Analysis and Pixel Spectrum Decomposition Method", which is listed in item 13 of the scope of patent application, where the genetic algorithm decomposition is composed of the following steps: -48- This paper scale is applicable to China National Standards (CNS) A4 Specifications (210X297mm) (Please read the notes on the back before filling this page}-Gutter (1) 一個編碼模塊隨機生成一組二進制字符串集合, 這址二進制字符串代表與該高譜圖象立方體的該像素^目 關的該S里估计,遺傳算法對這組二進制字符串進行操 作,且該組二進制字符串被送到一個解碼模塊者; 5 該解碼模塊對這紕二進制字符串進行解碼,該 解碼模塊的輪its g卩是藝該超級賴立雜像素點的 2含量估計集合,這組含量估計集合被送到一個適合度計 算模塊,該含量估計給出了該像素點中所包含的每個感興 趣材料的百分比; 10 (3)適合度計算模塊計算每個含量估計的適合度值, 在該適δ度什具模境令,指標函數取該二進制字符串(也 稱之為染色體)’並返回一個值,然後把該指標函數的值 映射為適合度以適應該遺傳算法,該適合度值是基於由該 字符串代表的所有可能解的性能的回答,該編碼字符串的 15該含量估計越好’該適合度值也越高,該適合度值再送到 一個繁殖模塊; (4)執行判別過程以確定是否終止進化,通過評價 該一進制字符串間的差別來進行判別,.當該字符串間的差 別小於一個預設值時,該遺傳算法退出進化,在此之後, 20選擇一個適合度值最大的二進制字符串作為解,其相應的 含量估計矢量就是該像素點的含量估計; (5)在該繁殖模塊中,基於來自該適合度計算模塊的 該適合度輸出,進行繁殖,在此,該繁殖是基於適者生存 的規律,即,這些適合度值高的二進制字符串會在新一代 -49-(1) An encoding module randomly generates a set of binary character strings. The binary character string represents the estimated value associated with the pixel of the hyperspectral image cube. The genetic algorithm operates on this set of binary character strings. And the set of binary strings is sent to a decoding module; 5 The decoding module decodes the binary string, and the round of the decoding module its g 卩 is a set of 2 content estimates for the super pixel This set of content estimation set is sent to a fitness calculation module, which gives the percentage of each material of interest contained in the pixel; 10 (3) The fitness calculation module calculates the content of each content estimation Fitness value. At the appropriate δ degree, if there is a model order, the indicator function takes the binary string (also called chromosome) 'and returns a value, and then maps the value of the indicator function to fitness to adapt to the genetic Algorithm, the fitness value is an answer based on the performance of all possible solutions represented by the string, the better the 15 content estimate of the encoded string is The higher the value, the fitness value is sent to a breeding module again. (4) A discrimination process is performed to determine whether to terminate the evolution, and the discrimination is made by evaluating the difference between the unary strings. When the difference between the strings is less than a pre- When the value is set, the genetic algorithm exits evolution. After that, 20 chooses a binary string with the highest fitness value as the solution, and its corresponding content estimation vector is the content estimation of the pixel; (5) in the breeding module Based on the fitness output from the fitness calculation module, breeding is performed. Here, the breeding is based on the law of survival of the fittest, that is, these binary strings with high fitness values will be in the new generation -49- 中請專利範Chinese patent 旦這些二補字符串被魏或複製作為 鋪在配對#合_ 即父配與變異,從而繁殖; 呆作 5 ⑷在交配模财,通過交換該二進制字符争(染 尾來形成子字符串集合,交配為該二 質,在;b鋪通過隨機過程來混合和匹配其期望的品 ;4;:中’首先’從該繁《塊形成的該匹配 ”二^出_新生成的字符_;其二,沿這兩 10 彳==擇-編蚊;其三,纖交換位置以 ⑺在變異魏t ’偶而改變在—個特定字符串位 =的值,該變異是對任何簡單位的怪久損失的一種保險 曰^該變異的發生概率極低,以至在字符串集合中平均 只有一個字符串發生變異;以及 15 、〜(δ)將該新的二進制字符串集合送到該解碼模塊, 然後按(2),(3),⑷,(5),⑹,⑺,以及⑻步驟執行者。 4〇如申請專利範@第17顧述之「高譜縣分析及像素點 頻譜分解綠」,其t該遺傳算法分解由以下步驟組成: 20 (1) 一個編碼模塊隨機生成而一組二進制字符串集 合,這組二進制字符串代表與該高譜圖象立方體的該像^ 點相關的該含量估計,遺傳算法對這組二進制字符串進行 操作,且該組二進制字符串被送到一個解碼模塊; (2) 該解碼模塊對這組二進制字符串進行解碼,該 解碼模塊的輸出即是關於該超級頻譜圖象立方體像素點的 秦 申請專利範園 在該算梅4蝴適合度值, 標函數取該二進制字符串(也 適應該遺傳算法:== 予付彳、表的所有可能解雜能的Θ答,_碼字符串的 10 15 經濟部智慧財產局員工消費合作社印製 該含量估計越好’該適合度值也越高,該適合度值再送到 一個繁殖模塊; (4) 執行判別過程以確定是否終止進化,在此,判 別指標;t義為總進化代數,當該遺傳算法遞推到該總進化 代數時,選擇其t-個具有最大適合度值的二進制字符串 作為解並且該遺鮮法退出進化,其相應的含量估計矢量 就是該像素點的含量估計; (5) 在該繁殖模塊中,基於來自該適合度計算模塊 的該適合度輸出,進行繁殖,在此,該繁殖是基於適者生 存的規律,即,這些適合度值高的二進制字符串會在新一 代中有大量複製,一旦這些二進制字:符争被繁殖或複製作 為下代使用,該一進制字符串將在配對集合中進行另外 兩類操作,即交配與變異,從而繁殖; (6) 在交配模塊中,通過交換該二進制字符串(染 色體)的頭和尾來形成子字符串集合,交配為該二進制字 符争提供一種機制通過隨機過程來混合和匹配其期望的品 -51-本紙張^AilL财關家料(CNS)A4規格( 210X297公^ 六 中靖專利範圍 隹j該賴過財,首先,贿繁顏塊形颜該匹配 =I選出兩個新生成的字符串;其二,沿這兩個字符串 2隨機選擇-败換位置;其三,交換該交換位置以 幾的所有字符; 以至在字符串集合中平均 ⑺異模财,偶岐變在—轉定字符串位 =的值。該變異是對任何醉位_久敏的—種保險 带施’該變異的發生概率極低, 裝 只有一個字符串發生變異;以及 js)將該新的二進制字符串集合送到該解碼模塊, 喊按(2) ’(3),(4),(5),⑹,⑺1及⑻步驟執行者。 訂 經濟部智慧財產局員工消費合作社印製 線丨· 張 紙 本 5Once these two complement strings are copied by Wei or copied as a pair # 合 _, that is, parental mating and mutation, thereby breeding; dumb 5 ⑷ in the mating model, the exchange of the binary character contention (dye tail to form a substring set , Mating is the second quality, in; b shop through a random process to mix and match its desired product; 4 ;: in the 'first' from the complex "the match formed by the block" two ^ out _ newly generated characters _; Secondly, along these two 10 彳 == 择-编 mosquito; Third, the fiber exchange position is changed by the 魏 魏 t t occasionally changed in the value of a specific string bit =, the mutation is strange to any simple bit A long-lost insurance: the probability of occurrence of the mutation is so low that only one string in the string set mutates on average; and 15, ~ (δ) sends the new binary string set to the decoding module, Then press (2), (3), ⑷, (5), ⑹, ⑺, and ⑻ step performers. 40. For example, “High-Spectrum County Analysis and Pixel Spectrum Decomposition Green” of the patent application @ 第 17 顾 述, Which t the genetic algorithm decomposition consists of the following steps: 20 (1) a The encoding module randomly generates a set of binary strings. The set of binary strings represents the content estimate related to the image point of the hyperspectral image cube. The genetic algorithm operates on the set of binary strings, and the group The binary string is sent to a decoding module; (2) The decoding module decodes the set of binary strings, and the output of the decoding module is the Qin application patent fan garden about the super-spectral image cube pixels. Mei 4 butterfly fitness value, the scale function takes the binary string (also adapted to the genetic algorithm: == prepaid, all possible solutions to the table's energy solution Θ answer, _ code string of 10 15 Intellectual Property Bureau of the Ministry of Economic Affairs The better the employee ’s consumer cooperative prints the content estimate, the higher the fitness value is, and the fitness value is sent to a breeding module; (4) The discrimination process is performed to determine whether to terminate the evolution. Here, the discrimination index; t means Total evolution algebra, when the genetic algorithm recurs to the total evolution algebra, select its t-binary strings with the largest fitness value as the solution and The legacy method exits evolution, and its corresponding content estimation vector is the content estimate of the pixel; (5) In the breeding module, breeding is performed based on the fitness output from the fitness calculation module, and here, the breeding It is based on the law of survival of the fittest, that is, these binary strings with high fitness values will be copied in a large number of new generations. Once these binary words: runes are reproduced or copied for use as the next generation, the binary string will be Perform two other types of operations in the pairing set, that is, mating and mutation, to reproduce; (6) In the mating module, exchange the head and tail of the binary string (chromosome) to form a substring set, and mate to the binary Character contention provides a mechanism to mix and match its desired product through a random process-51-this paper ^ AilL Financial Relations (CNS) A4 specification (210X297 public ^ Liu Zhongjing patent scope) To match the block face, the matching = I select two newly generated strings; second, randomly select the two strings along the two-the position of failure; third, exchange the exchange All characters in a set of several; ⑺ average differential mode as well as in the set of character strings Choi, Qi becomes even in - rpm = value given bit string. The mutation is applied to any drunk bitch—a kind of safety belt, the probability of occurrence of the mutation is extremely low, and only one string is mutated; and js) sends the new binary string set to the decoding module, shouts Press (2) '(3), (4), (5), ⑹, ⑺1 and ⑻ step performers. Printed by the Consumer Property Cooperative of the Intellectual Property Bureau of the Ministry of Economic Affairs 丨 · Sheet 5
TW089108706A 1999-05-05 2000-05-04 Method for hyperspectral imagery exploitation and pixel spectral unmixing TW503375B (en)

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
US09/351,349 US6665438B1 (en) 1999-05-05 1999-05-05 Method for hyperspectral imagery exploitation and pixel spectral unmixing

Publications (1)

Publication Number Publication Date
TW503375B true TW503375B (en) 2002-09-21

Family

ID=27613488

Family Applications (1)

Application Number Title Priority Date Filing Date
TW089108706A TW503375B (en) 1999-05-05 2000-05-04 Method for hyperspectral imagery exploitation and pixel spectral unmixing

Country Status (1)

Country Link
TW (1) TW503375B (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103822711A (en) * 2014-03-03 2014-05-28 中国科学院遥感与数字地球研究所 Digital image display method and hyper-spectral telescope
CN104268561A (en) * 2014-09-15 2015-01-07 西安电子科技大学 Hyperspectral image mixing eliminating method based on structure prior low rank representation
CN108280396B (en) * 2017-12-25 2020-04-14 西安电子科技大学 Hyperspectral image classification method based on depth multi-feature active migration network

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103822711A (en) * 2014-03-03 2014-05-28 中国科学院遥感与数字地球研究所 Digital image display method and hyper-spectral telescope
CN103822711B (en) * 2014-03-03 2015-12-02 中国科学院遥感与数字地球研究所 Digital image display methods and EO-1 hyperion telescope
CN104268561A (en) * 2014-09-15 2015-01-07 西安电子科技大学 Hyperspectral image mixing eliminating method based on structure prior low rank representation
CN104268561B (en) * 2014-09-15 2017-08-25 西安电子科技大学 High spectrum image solution mixing method based on structure priori low-rank representation
CN108280396B (en) * 2017-12-25 2020-04-14 西安电子科技大学 Hyperspectral image classification method based on depth multi-feature active migration network

Similar Documents

Publication Publication Date Title
Ru et al. Multi-temporal scene classification and scene change detection with correlation based fusion
CN109522857B (en) People number estimation method based on generation type confrontation network model
St-Yves et al. Generative adversarial networks conditioned on brain activity reconstruct seen images
CN106023065B (en) A kind of tensor type high spectrum image spectral-spatial dimension reduction method based on depth convolutional neural networks
Li et al. Using wavelet transform and multi-class least square support vector machine in multi-spectral imaging classification of Chinese famous tea
US6665438B1 (en) Method for hyperspectral imagery exploitation and pixel spectral unmixing
Valentin et al. Categorization and identification of human face images by neural networks: A review of the linear autoassociative and principal component approaches
CN108875459B (en) Weighting sparse representation face recognition method and system based on sparse coefficient similarity
CN110728629A (en) Image set enhancement method for resisting attack
CN114548428B (en) Intelligent attack detection method and device of federated learning model based on instance reconstruction
Zhang et al. IL-GAN: Illumination-invariant representation learning for single sample face recognition
Masood et al. Differential evolution based advised SVM for histopathalogical image analysis for skin cancer detection
CN116843400A (en) Block chain carbon emission transaction anomaly detection method and device based on graph representation learning
CN110175642A (en) A kind of chrysanthemum similarity calculation method based on PCA dimensionality reduction and feature binary
CN113033305A (en) Living body detection method, living body detection device, terminal equipment and storage medium
CN118365192A (en) Water environment quality remote sensing analysis method, system, electronic equipment and storage medium
CN112541530B (en) Data preprocessing method and device for clustering model
TW503375B (en) Method for hyperspectral imagery exploitation and pixel spectral unmixing
CN114596464A (en) Multi-feature interactive unsupervised target detection method and system, electronic device and readable storage medium
CN116993839B (en) Coding mode screening method and device, electronic equipment and storage medium
CN117671800A (en) Human body posture estimation method and device for shielding and electronic equipment
Audhkhasi et al. Data-dependent evaluator modeling and its application to emotional valence classification from speech.
Dhar et al. Distill and de-bias: Mitigating bias in face verification using knowledge distillation
Liu et al. Discriminative face hallucination via locality-constrained and category embedding representation
CN113792541B (en) Aspect-level emotion analysis method introducing mutual information regularizer

Legal Events

Date Code Title Description
GD4A Issue of patent certificate for granted invention patent
MM4A Annulment or lapse of patent due to non-payment of fees