201044877 「、發明說明: 【發明所屬之技術領域】 本發明為有關於-種影像監視設備 影像解析自動調適之方法,尤豆是浐處理糸統及其 产的,It报τ 其疋彳日一種在任何環境光線明暗 度的跡下’皆能核影像解析之㈣及方法。 【先前技術】 Ο ❹ 發生^事^於=的=,能夠以連續方式紀錄某段時間所 ίί事因此,磁用來❹校通路況或監視住家、公司 turr知道路f住家、公司目前所處之狀況。而動態 處織處理後,於顯示裝置顯示。然而,: 置顯示之猶仍有喊,無法清_^'^。而5在顯不裝 -雜ίί案文巧中,透過傳統CCD錄出視頻訊號,並運用 間摺積料付法與_航演算程式分析斷層掃描 貝枓=⑽較仰方法之反演料彡像解析度 US7444QG3的驗證麵,皆細重點隱高解析 飽和層模型。在簡單陣列感測器的成像模“ η料之景彡像與運雜式之準確度相當,諸如傳統『直方 ,』、馬卡夫法』或『邊界檢測』為基礎等方法。在穩鎮盥暫 υ模擬試驗中’比較兩種不同演算法,_衰減寅算 =像,析度’通常減速反影像要差,這齡實際情況 了二雷達波傳遞可能有額外之能量損失,因此須於反 程中加以考量。 在暫態的試驗中’尚斯傅利葉濾波反演算影像差值顯示, 阻尼最小平方法採用阻尼,以保證反演算之順利進行,便能減 201044877 要較傳ΐϋί巧值與最低侧之誤差程度,但在解析度上 介,阳鬲通濾波器來的差。若比較所得之衰減率變 更接进平方法所得之值,卻較傳統CCD高通濾波器, 定值,此顯示採雜尼最小平方法所取得之 ί這合ϋΐϊίCCD高通濾波器更接近模型原始設定值, 須依’因此欲在解析度與準確度間得到平衡,必 紐’以制最佳彳b之解減鮮確度。 【發明内容】 Ο Ο 本發明為有關於一種可將CCD擷取進來的視訊籍禮卢 ^此^昇影像的解析度;在處理影像解析度提昇之前,ΐ 二要的前處理程序需要進行,這些前處理壞有 ‘取進ίίίί像解析的品f,因此本發明針對CCD視 廓切月斤度影像前處理的方法包括影像分割與輪 上的i长ίίϊ疋—個非常困難的技術’尤其對於㈣應S 、处由於为割演算法需要龐大計算量,因此在計曾旦 法二確的改良’便可成功避開傳統影像 ,達到快速完成切割之要求,以讓整個ΐί 處理單L P〇Wer【低運算量】的CPU【令央 】上元成影像解析的動作。 、 本發明為_—種高解析度還原的方法,採 2,亚將人類眼睛的視覺特性加入影像分割之象 出彻析度影像還原的核心本發明·201044877 ", the invention description: [Technical field of the invention] The present invention relates to a method for automatically adjusting image analysis of an image monitoring device, and the bean is processed by the cockroach and its production. Under the trace of any ambient light and darkness, it can be used to analyze the image (4) and methods. [Previous technique] Ο ❹ The occurrence of ^^^== can record a certain period of time in a continuous manner. Come to the school to observe the situation or monitor the home, the company turr knows the road f home, the company's current situation. After the dynamic processing, it is displayed on the display device. However, the display still shouts, can not clear _^ '^. And 5 in the display is not installed - mis - ίί text in the text, through the traditional CCD recording video signal, and using the inter-folded accumulation method and _ aeronautical calculation program to analyze the tomography shellfish = (10) inversion method The verification surface of the image resolution US7444QG3 is focused on the hidden high-resolution saturated layer model. The imaging model of the simple array sensor is equivalent to the accuracy of the image, such as the traditional “straight square,” Makaf "Or" edge detection "is based on other methods. In the stable town 盥 temporary simulation test, 'comparing two different algorithms, _ attenuation = = image, resolution 'is usually slow down the inverse image, the actual situation of the second radar wave transmission may have additional energy loss, Therefore, it must be considered in the reverse process. In the transient test, the Shangs Fourier filter inversion calculation image difference display shows that the damping least square method uses damping to ensure that the inversion calculation is carried out smoothly, and the degree of error between the 值 巧 value and the lowest side can be reduced by 201044877. However, in terms of resolution, the difference between the Yangshuo pass filter. If the comparison of the obtained attenuation rate is changed to the value obtained by the flat method, it is more fixed than the traditional CCD high-pass filter. This shows that the CCD CCD high-pass filter is closer to the original set value of the model. , according to 'therefore want to balance between resolution and accuracy, Bian's solution to the best 彳b solution. SUMMARY OF THE INVENTION The present invention relates to a resolution of a video library that can capture a CCD, and a pre-processing procedure that needs to be performed before processing image resolution is improved. These pre-processing methods are 'fetched into the ίίίί image, so the method for pre-processing the CCD view of the CCD profile includes image segmentation and i-long on the wheel—a very difficult technique. For (4) should be S, because the cutting algorithm requires a large amount of calculation, so in the improvement of the Zeng Dan method, you can successfully avoid the traditional image, and achieve the requirement of quickly completing the cutting, so that the whole ΐί processing single LP〇 Wer [low computation] CPU [Command] is the action of the image analysis. The present invention is a high-resolution reduction method, and the second embodiment of the human eye is added to the image segmentation image.
Bounda㈣版咖)的人類視覺暫留 題界£域_方法,以解決演算法會有過度切= 201044877 =·將原始影像簡取_錢理崎析度,触a」β =則,以發展出細多重解析度的精確影像“之前處王】 出JBD七刀割完祕的内部影像内插多重指積之方法 切割後的演算法之邊界區域内的影像做解析度的提昇。將 覺特㈣勸輯則規職丨ANCE與 【實施方式】 ❹ Ο 露,容'目的及功效有更清楚及完整的揭 路特舉只鞑例,亚配合所附圖式,詳細說明如下: 本發明影像監視設備之影像解析自動 度料之前,彡魏進行高解析度影^乍理 而解析度影像前處理的方法包括影像分割與輪 切割’其實現之方法至少包括: ^以具有人類視覺仙特性的細(」usi B()unda「y Diffe 邊界區域内插方法,進行影像切割; 2. 將經過JBD(Just Boundary Different)邊界區域内插方法處 理過之影像,進行」BD乡重㈣度的料影像蝴之前處理; 3. 在JBD夕重解析度的精確影像切割完成後,以内部影像内 插多重摺積之方法,將蝴後的演算法之邊界區域 解析度的提昇; 4. 最後,透結合視覺特性的影像評估準則JBD_VAmANcE與 JBD_SNR完成影像解析自動調適。 八 而達成上述本發明之影像監視設備的影像解析自動調適 方法所给之糸統(以下請參第一圖),包括: 至少一動態影像輸入設備(1 ),以即時擷取動態影像,· 201044877 -邊界偵測模組(2 ),具有至少—内插濾波器(2工) 及一影像強度遮罩矩陣(2 2 ),係用以將由_影_入設 備⑴輸人之_影像巾,切_所需要高解析度還原晝面 之邊界,以供後續濾波細插之處理,且該邊界酬模組⑵ 以其本身切割時間為準; 问解析度内插模組(3 ),輕接於該動態影像輸入設備 ⑵與影像輸出設備⑷之間,係負責在每4個像素點中Bounda (four) version of the human vision of the human visual persistence problem domain _ method to solve the algorithm will have excessive cut = 201044877 = · the original image is taken _ Qian Liqi resolution, touch a "β = then to develop Accurate image of fine multiple resolution "Before the King" Out of the JBD seven-knife cut the internal image interpolation multi-finger method The image in the boundary area of the algorithm after cutting is improved. Persuasion is the standard 丨 与 【 【 【 【 , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , Before the image analysis of the device is automatically calculated, the method of high-resolution image processing and resolution image pre-processing includes image segmentation and wheel cutting. The method for achieving the method includes at least: ("usi B()unda" y Diffe boundary area interpolation method for image cutting; 2. Image processed by JBD (Just Boundary Different) boundary area interpolation method, "BD" weight (four) degree image Before processing 3. After the accurate image cutting of the JDF re-resolution is completed, the resolution of the boundary region of the algorithm after the butterfly is improved by interpolating the multiple images of the internal image; 4. Finally, the visual characteristics are combined. The image evaluation criteria JBD_VAmANcE and JBD_SNR are automatically adapted to the image analysis. The image analysis device of the present invention is provided with the image analysis automatic adjustment method (hereinafter referred to as the first figure), including: at least one dynamic image input device (1), to capture dynamic images in real time, · 201044877 - Boundary detection module (2), with at least - interpolation filter (2 work) and an image intensity mask matrix (2 2), used to be _影_入设备(1) Input _image towel, cut _ requires high resolution to restore the boundary of the 昼 surface for subsequent filtering and fine interpolation, and the boundary compensation module (2) is subject to its own cutting time; The resolution interpolation module (3) is lightly connected between the dynamic image input device (2) and the image output device (4), and is responsible for every 4 pixels.
内插出1個新的像素點,再經由平坦傾波器將高解析度影像 的尖銳邊界平坦化; 至〉、-影像輸出設備(4 ),係供顯示處理後之影像。 像切高解析度影爾理達成精物 覺暫留特性的邊界區域内插方法,以解決: 統凟异法會有過度切割之問題。 ί 2移生/時,;BD演算法的邊界高頻濾波11便 應^ ’並與,個視訊_縣值微分運算 部份辦财,並由新產生的魏點所引發 2ίί «便成夠在腦演算法中,將非邊界的部份排除, 誤料致制高崎妓像處_處理到不 驟,這將會直接影響到將來ΐ嵌 =fwCPU會造摘外犧 般而言,勘可區別人類眼睛邊界輪軸能力,終由此 此力將可_景或背景_練對比度有___:因此 201044877 針對JBD的實驗,如第二圖所示,可讓測試者看著—個平整 均勻照度的圖案影像框(1 〇 〇 ),在區域中心產生—個矩形 的方塊,其亮度強度(1 〇 1 )為Ρ+ΔΡ。 夕 本發明所引用的JBD模型如下: JBD(s(x,y)) = ^bcg(x,y) - 255 for g(x,y) >255 •/5D(g(x,_y))二士x(3 -f〇rg(x,y) ^255 其中g(x,y)為一張影像在(X,y)位置像素點之亮度值,將該 值送入JBD(g(x,y))函式為此像素點,以求出其對應之JBd值, ❹ 在JBE)函式中7 = 1.7J = 0_873。由於JBD值的檢測可得知, 人類眼睛對於高亮度的敏感度較佳,反之對低亮度的敏感度較 差。 〜 ^由於一張影像可能會包含很多不必要的雜訊,這些雜訊通 f會影響到往後影像邊界切割的動作,導致無法簡潔有效的表 示出原影像,這將對於後續的内插運算造成極大的負荷,使得 • 在Low Computing Power的嵌入式系統的性能大為下降。本發 明將,用平坦化來解決上述之問題,平坦化基本上是一個低^ . 濾波器,它主要用來使影像消除模糊或降低雜訊。本發明設計 〇 一個5x5的平坦化濾波器之遮罩,如第三圖所示,該平坦彳°匕濾 波器包含一個二維矩陣(丄〇 2 ),二維矩陣(丄0 2 )裏^ 的7L素是1 ’外面則是乘以5分之i。通常越大的遮罩模糊效 果越強,相當於此濾波器的截止頻率越來越低,高頻部份被過 遽、掉的越多。 、在本實施例中,JBD多重解析度的精確影像切割之前處理 方法,如第四圖所示,由多重JBD所組成jgD群組(丄丄〇), 在每一组腿群組(1 10)的輸出串接一個JBD高通濾、波 1 1)’以將有變化的邊界過濾出來,最後再將所有JBD 南通濾'波器(1 1 1)的輸出經由影像框相關器【ImageFrame orrelation】(1 1 2 )整合出來’即可得多重景多像切割的圖案 201044877 或動晝之邊界。 -般而言」發展-般化容忍變形延長在c。她t ,】的相互關係猶的設計過財,早 :計,係以被規定的模式對平面旋轉過程與在影像】 中的Image Frame【影像框】強度,此為改變處】 Ο 項挑碱。本發明將發展一種新保護訊息的機率影 圖像和多個過濾器合併恤關係產量。機忿 互關係產量之空間和暫時關係,這種機率影像框,才目 4識別與相反跟蹤,再應_神經網路方法作 】 之辨識,進而求得各影像之瞬間變動。 貝2運動 〇 為瞭解類神綱路方法之優、缺點,魏影像亦 相關【^祕如—⑽】方式來進行分析與比較,以 ;速度場,結果顯示類神經網路方法之計算速 妨式為快,且f點之辨識率也較傳統方式好。者 f動時’若經由霍夫轉換【HoughTransform】後:在辨 發生_,這將會造成辨識誤差,因此在計算過 Ί聲時’將會作—次角度改變,即是將質點運動斜率在5 =2。之間轉換成36〇。〜曹’便可歧質點接近 ^斜率落在細象關第—象限無法辨識的問題。在 程度時,由於部份質點已辨識出,剩餘的質點亦不°多了护 :便可將所有繼質點再作45。的座標轉換, 質點之分類_。 直運動 田衫像中質點被分類出來後,便可以繼續辨 標位置,而計算質點的水平⑻與垂直速度(V)公式^下座 txxuAt = 12 tyyvAt=l2 ,、中(xl,yl)為第一張影像質點之χ、y座標,而 ”'、(X , yl)下—瞬間之座標值,At為兩影像之拍攝時間之間隔 201044877 (在此為〇·〇〇4秒)。 切到割完成後的内部影像内插多重摺積之方法,將 一 ^ ί的异去之邊界區域内的影像做解析度的提昇。針對每 H彡,’首先依照影像的_邊界濾波值由高至低排列, 亚儲存在二維陣列的記憶體内,且必需狀下列條件·· (Α)針對-張影像的像素做直接或間接存取。 (gf」個'在處理-的視訊框’亦能對鄰近的像素點進行内插 扞、軍瞀二所示的㈣切割後後影像内插運算,JBD内 素點ΤΤίρ、p3 Q)為内插運算範圍之區塊,内插出來的像 二、d點笔ϋ點為内插參考像素點(1 3 1 ) a點、b ,點、 插運算=_」_插運频組(1 3⑻中經由内 在=圖中提出調節錄像的基本算法,本發明使用了「增 ΟA new pixel is interpolated, and the sharp boundary of the high-resolution image is flattened by a flat dipper; to the >, - image output device (4) is used to display the processed image. The interpolation method of the boundary region is achieved by cutting the high resolution image to achieve the persistence characteristic of the essence, so as to solve the problem of over-cutting. ί 2 transfer / time,; BD algorithm boundary high-frequency filtering 11 should be ^ 'and with a video _ county value differential operation part of the fortune, and caused by the newly generated Wei point 2 ίί « In the brain algorithm, the non-boundary part is excluded, and the result is that the processing of the Takasaki image is not processed, which will directly affect the future embedded=fwCPU will be smashed. Differentiating the human eye's boundary wheel axle ability, the force will be __ or background_practice contrast ___: therefore 201044877 for JBD experiments, as shown in the second figure, allows the tester to look at - a flat uniform illumination The pattern image frame (1 〇〇) produces a rectangular square at the center of the area with a brightness intensity (1 〇 1 ) of Ρ + ΔΡ. The JBD model cited in the present invention is as follows: JBD(s(x,y)) = ^bcg(x,y) - 255 for g(x,y) >255 •/5D(g(x,_y))二士x(3 -f〇rg(x,y) ^255 where g(x,y) is the luminance value of a pixel at the (X,y) position, and the value is sent to JBD (g(x , y)) The function is the pixel to find its corresponding JBd value, 7 in the JBE) function 7 = 1.7J = 0_873. As the JBD value is detected, it is known that the human eye is more sensitive to high brightness, and vice versa. ~ ^ Since an image may contain a lot of unnecessary noise, these noises will affect the cutting of the image boundary in the future, resulting in the inability to display the original image in a concise and effective manner, which will be used for subsequent interpolation operations. The enormous load caused the performance of the embedded system in Low Computing Power to drop dramatically. The present invention will solve the above problem by flattening. The flattening is basically a low filter, which is mainly used to eliminate image blur or reduce noise. The present invention is designed as a mask of a 5x5 flattening filter. As shown in the third figure, the flat 彳°匕 filter comprises a two-dimensional matrix (丄〇2) in a two-dimensional matrix (丄0 2 ). ^ 7L is 1 'outside is multiplied by 5 points. Generally, the larger the mask blur effect, the lower the cutoff frequency of the filter, and the higher the high frequency portion is over and over. In this embodiment, the JBD multi-resolution accurate image cutting pre-processing method, as shown in the fourth figure, consists of multiple JBDs composed of jgD groups (丄丄〇), in each group of leg groups (1 10 The output of the JBD is connected in series with a JBD high-pass filter, wave 1 1)' to filter out the changed boundary, and finally the output of all JBD Nantong filter (1 1 1) is passed through the image frame correlator [ImageFrame orrelation] (1 1 2) Integrate it's much more than a multi-faceted cut pattern 201044877 or the boundary of the moving edge. - Generally speaking, the development-generalized tolerance deformation is extended in c. Her t, the relationship between the two is still designed too rich, early: count, according to the specified mode of the plane rotation process and the image in the image] [Image Frame] intensity, this is the change] Ο . The present invention will develop a new generation of protection messages with a probability image and multiple filter conjugate relationship yields. The spatial and temporal relationship between the production and the relationship of the machine, the probability frame, the identification and the opposite tracking, and then the identification of the neural network method, and then the instantaneous changes of each image. In order to understand the advantages and disadvantages of the method of class-like roads, Wei 2 is also related to the method of [^ secret--(10)] for analysis and comparison, and the speed field shows the calculation of the neural network method. The formula is fast, and the recognition rate of the f point is also better than the conventional method. When f is moving, 'If HoughTransform is converted via Hough: _, _, which will cause identification error, so when calculating the humming sound, 'will make the angle change, that is, the slope of the particle motion is 5 = 2. Convert between 36〇. ~ Cao's can be close to the mass point. ^ The slope falls on the fine-grain--the quadrant is unrecognizable. At the same time, since some of the particles have been identified, the remaining particles are not much more protected: all the successive points can be further 45. Coordinate conversion, classification of particles _. After the straight-moving shirts are sorted out like the medium-mass points, they can continue to identify the position, and calculate the level of the mass point (8) and the vertical speed (V) formula ^ the lower seat txxuAt = 12 tyyvAt = l2 , and the middle (xl, yl) is The first image is the mass point, the y coordinate, and the '', (X, yl) is the instantaneous coordinate value, and At is the interval between the shooting times of the two images 201044877 (here, 〇·〇〇 4 seconds). The method of interpolating multiple folds into the internal image after the cut is completed, and the resolution of the image in the boundary region of the image is improved. For each H彡, 'the filter value of the image is firstly increased according to the image. Arranged to the lowest, sub-stored in the memory of the two-dimensional array, and the following conditions are necessary: (Α) Direct or indirect access to the pixels of the image (gf" in the processing - video frame It is also possible to interpolate adjacent pixels, and (4) the image interpolation operation after cutting (4), and the JBD inner point ΤΤίρ, p3 Q) is the block of the interpolation operation range, and the interpolated image Second, the d point of the pen point is the interpolated reference pixel (1 3 1) a point, b, point, insert operation = _" _ insert frequency In the group (1 3 (8), the basic algorithm for adjusting the video is proposed via the internal = map, and the present invention uses the "increase"
Si適合做運算的内插演算法,以下介紹增強型 增強型内插演算法: 項的將放大/縮小的比例代人特定二次 表現較好,但可能點^ °對於低頻訊號較多的影像 也合比料祕、上生運"'娜度較高之缺點,因此處理速度 像;。1 2 1 =;去慢。增強型内插演算法可透過4個原始 像京點(121)内插運算出一個新 平i之基本單位:當開始同時處理水 採時,便開始 圖增強型_算法執行流程圖點丄12 $由第, 這個階段内,每-組美太_鮮:f]—者疋叫並仃的。在 區之内插演算,使内;2 2個雏1 2〇)進行2x2維度 邛2個'准度區的類型邊界被確定。當水 9 Ο ❹ 201044877 二=冗直環 流程圖,水平同步判別模电j:強型像素内插演算法執行 直同步判別模組則代插運算,而在垂 的像素點。 内插運异,運鼻完成後產生新 其t,第七圖的運算流程如下 1 CT 1 \ 理解析度,·原轉像财取_所㈣水平同步處 步驟二(]C; 9、. 4 理解析度,· 字原始影像做次取樣到所要的垂直同步處 前-個影勵—S】 〇LDH—口與0咖c .】灼水平成份的失真值【Distortion】 盥目前男Hi)’邊界取樣並設定出前一個影像框【Previous】 與目〇=,【Cu咖t】的垂直成份的Dist〇rti〇n值〇LDv_p 步驟五(1 5 q.屯丨V丨α 如果全邻初判別刖的影像框所有像素是否都取出來, 右目^ϊΐ出來則不會得到新的像素點,且因此結束’如果 有’則跳到步驟六進行内插算; 1 5 6 ):進行增強型内插法之内插運算,目的是為 了内插出-個新的像素於目前的影像框。 σ視覺特性的影像評估準則JBD-VARIANCE與 以下式定義本發明的JBD_VARIANCE : 10 201044877 JBD Variance ΣΝΚ ΓΓ| Κ=ι l^jn J JJ xj.k JBD(xk) χδ j,k (公式一) δ Μ iif xJ,k ~ Xk >JBD(xk) 其中m為第k個區域内所有像素的邊界亮彦值,而- $ 内所有躲點的平均亮度值,咖(A )為 均 的焭度值&之臨界JBD值。 一般而言,JBD—VARIANCE 與 JBD—PSNR 皆需者;f Ο Ο ίΓ性改但,-从齡臓為利用具有多元性的影“割準 則’以強调影像分割區域’因為要配合内插演算法,必 ,,區塊與多重内插節點的結合。本發明考慮在同 域的母個像素點與前-麵赌的互相_ 得 -個變異度的差異值。藉由這些差異值進仃以= JBD_PSNR tPeak Signal to Noise Ratio ; ί ’。掉的雜訊,並轉‘始訊 -般而言,最鄰近均勻化法可_為助 ==T【:】t: 同時利用影像灰階值^ 作方ΐ通'^ ί、=呆留影像特徵與平滑雜訊。操 由左至右與由上而下來對影像進行過^以) t像亡 -對稱像素元組中灰階值,_用最^ 内母 計算平均值或中位數,以取代核吻中二象,(义階值’ 種方法的優點,在於濾除雜訊的同時^^階,採用此 邊界資訊的反差。 奸保持賴物屋角與 請參第八圖,為本發明所提出的7 (1 4 0 )【Filtei·】,其操作朗於下: D—SNR過濾器 11 201044877 在灰階方框像素元(141)内可取出3組對稱像素 90、90、95 ; 90、90、95 ; 103、1〇3、108 ;將各組像素元中 之灰階值與中心像素元之灰階值90相比較後,於各植中取出 其中一組最接近該中心像素元灰階值90之灰階值,其分別 90、90與95,將所取得之3個灰階值予以平均得到91'D並 以此數值取代原始中心像素元灰階值90。 、’ 在彩度方框像素元(1 4 2)為5x5的方陣,内可取出5組 Ο Ο 對稱像素元:90、95、104、103、103 ; 1〇3、1〇8、115、115、 116 ; 115、124、129、130、132 ; 126、136、14卜 14心 145 | 138、—143、152、151、155 ;將各組像素元中之灰階值鱼中心 像素元之灰階值129相比較後,於各組中取出其中一袓^接近 該中心像素元灰階值129之彩度值,其分別為9〇、%、丨⑽3; 103與103,將所取得之5個彩度值予以平均得到99,並以此 數值取代原始_心像素元灰階值129。接著再利用(公式 出腦―VARIANCE的公式,以得到梯度大小和方向後,再依 ,度方向對梯度大小作非最大值刪除(跑—施如職 卩在梯度方肖上邊界點像錢大小值,應該大 於u像素的A小值,因此只取區域(LQeai)最大值為邊界 ^再只施一個附加遲滯性界s(Hysteresis Thresh〇lding)的步 Ϊΐ删除不正確的邊界點’並採用兩個臨界值,一個為高臨 另一個為低臨界值尽。任何一個像素的大小值,只 兒士 貝1J可指定其為邊界點’而連接此點的像素,只要 ^小值大於心,亦可被指定為SNR【訊號/雜訊比】點。 用以二戶f述’ _本發明以—實施例揭錄如上,然並非 在不。本發明所屬技術領域巾具有通常知識者, 闵t卜太月,精神和範圍内’當可作各種之更動與濶飾, *明之保護範圍,當視後附之申請專利範圍所界定者為 圖式簡單說明】 12 201044877 第一圖:顯示根據本發明實施例所需之系統示意圖 第二圖:顯示根據本發明實施例之JBD邊界檢測示意圖 第三圖:顯示根據本發明實施例的平坦化濾波器 第四圖:顯示二維度基本像素框内插 第五圖:JBD多重解析之精確影像切割示意圖 第六圖:JBD内插運算示意圖 第七圖:增強型像素内插法示意圖 0 第八圖:一個7 x7 JBD_SNR過濾器 【主要元件符號說明】Si is suitable for the interpolation algorithm of the operation. The following describes the enhanced enhanced interpolation algorithm: The ratio of the enlargement/reduction of the item is better for the specific secondary performance, but it may be for the image with more low frequency signals. It also has the disadvantages of higher quality than the secret of the material, and the higher the speed of the Nadu. 1 2 1 =; go slow. The enhanced interpolation algorithm can calculate the basic unit of a new flat i by interpolating four original objects like Jingdian (121): when starting to process water extraction at the same time, it starts to enhance the graph _ algorithm execution flow chart point 丄12 $ By the first, in this stage, each group is too _ fresh: f] - the screaming and screaming. Insert the calculation within the zone, make the inner; 2 2 chicks 1 2〇) carry out the 2x2 dimension 邛 2 'precision zone type boundaries are determined. When water 9 Ο ❹ 201044877 2 = redundant loop flow chart, horizontal synchronization discriminant mode j: strong pixel interpolation algorithm execution The straight synchronization discriminating module interpolates the operation while at the pixel point. The interpolation process is different, and the new nose is generated after the nose is completed. The operation flow of the seventh figure is as follows: 1 CT 1 \ Rational resolution, · The original conversion image is taken _ (4) Horizontal synchronization step 2 (] C; 9,. 4 resolution, · the original image of the word is sampled to the desired vertical synchronization before - a shadow excitation - S] 〇 LDH - mouth and 0 coffee c.] distortion level component distortion value [Distortion] 盥 current male Hi) 'Boundary sampling and setting the previous image frame [Previous] with the target =, [Cu coffee t] vertical component of the Dist〇rti〇n value 〇 LDv_p step five (1 5 q. 屯丨 V 丨 α if the full neighbor It is determined whether all pixels of the image frame of the frame are taken out, and the new pixel point will not be obtained when the right eye is clicked, and thus the end of 'if there is' then jump to step six for interpolation; 1 5 6 ): perform enhanced type The interpolation operation of the interpolation method is to insert a new pixel into the current image frame. The image evaluation criterion JBD-VARIANCE of the σ visual characteristic defines the JBD_VARIANCE of the present invention with the following formula: 10 201044877 JBD Variance ΣΝΚ ΓΓ| Κ=ι l^jn J JJ xj.k JBD(xk) χδ j,k (Formula 1) δ Μ iif xJ,k ~ Xk >JBD(xk) where m is the boundary brightness value of all pixels in the kth region, and the average brightness value of all hidden points in - $, coffee (A) is the average value of the mean & critical JBD value. In general, JBD-VARIANCE and JBD-PSNR are required; f Ο Ο Γ Γ 改 改 , , - - - - - - - - 臓 臓 臓 臓 利用 利用 利用 利用 利用 利用 利用 利用 利用 利用 利用 利用 利用 利用 利用 利用 利用 利用 利用 利用 利用 因为 因为 因为 因为 因为 因为The algorithm, the bound, the combination of the block and the multiple interpolation nodes. The present invention considers the difference between the mutual _ _ variability of the parent pixel and the front-surface gambling in the same domain. = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = Value ^ ΐ ΐ ' '^ ί, = stay image features and smooth noise. Operate from left to right and from top to bottom on the image ^) t-like-symmetric pixel tuple grayscale value, _ Calculate the mean or median with the innermost mother to replace the second image in the nuclear kiss. The advantage of the method is that the noise is filtered out and the contrast of the boundary information is used. Rape keeps the object and asks for the eighth picture, which is 7 (1 4 0 ) [Filtei·] proposed by the present invention, and its operation is as follows: D— SNR filter 11 201044877 Three sets of symmetric pixels 90, 90, 95 can be taken out in the gray-scale box pixel (141); 90, 90, 95; 103, 1〇3, 108; gray in each group of pixel elements After comparing the order value with the gray scale value 90 of the central pixel element, one set of gray scale values closest to the center pixel element gray scale value 90, which are respectively 90, 90 and 95, are taken out from each plant. The three grayscale values are averaged to obtain 91'D and the original center pixel grayscale value is replaced by this value. 90. 'In the chroma box, the pixel (1 4 2) is a 5x5 square matrix, and 5 groups can be taken out.对称 Symmetrical pixel elements: 90, 95, 104, 103, 103; 1〇3, 1〇8, 115, 115, 116; 115, 124, 129, 130, 132; 126, 136, 14b 14 heart 145 | 138 ——, 143, 152, 151, 155; comparing the grayscale value 129 of the grayscale value fish center pixel element in each group of pixel elements, taking one of the groups to obtain the grayscale value of the central pixel element The chroma values of 129 are 9〇, %, 丨(10)3; 103 and 103, and the obtained 5 chroma values are averaged to obtain 99, and the original _ heart pixel gray scale is replaced by this value. 129. Then reuse (form the brain - VARIANCE formula to get the gradient size and direction, and then according to the degree direction to the gradient size for the non-maximum deletion (run - Shi Shizhan on the gradient square image on the boundary point image The value of the money should be larger than the A value of the u pixel, so only the maximum value of the region (LQeai) is taken as the boundary. Then, only the step of the Hysteresis Thresh〇lding is added to delete the incorrect boundary point. And use two critical values, one for the high and the other for the low threshold. For any pixel size value, only the pixel can be specified as the boundary point and the pixel connected to this point can be specified as the SNR [signal/noise ratio] point as long as the small value is greater than the heart. For the two households, the invention is disclosed above, but not the same. The technical field of the invention belongs to the general knowledge, 闵t Bu Taiyue, spirit and scope 'when it can be used for various changes and decoration, * the scope of protection, as defined by the scope of the patent application Brief Description of the Invention] 12 201044877 First Figure: A schematic diagram showing a system required according to an embodiment of the present invention. FIG. 2 is a third schematic view showing JBD boundary detection according to an embodiment of the present invention: showing flattening filtering according to an embodiment of the present invention. The fourth picture shows the two-dimensional basic pixel frame interpolation. The fifth picture: JBD multi-resolution accurate image cutting diagram. The sixth picture: JBD interpolation operation diagram. The seventh picture: enhanced pixel interpolation method diagram 0. A 7 x7 JBD_SNR filter [main component symbol description]
動態影像輸入設備 邊界偵測模組 内插濾波器 影像強度遮罩矩陣 高解析度内插模組 影像輸出設備 (10 0)影像框 (101)亮度強度 (10 2)二維矩陣 (1 1〇)JBD群組 (1 1 1 ) JBD高通濾波器 (112)影像框相關器 (1 3 0 ) JBD内插運算模組 201044877 (13 2)内插出來的像素點 (131)内插參考像素點 (121)原始像素點 (12 2)新的像素點 (12 0)二維像素框 (12 4)水平環型佇列 (1 2 3)垂直環型佇列 〇 ( 1 4 ◦ ) JBD—SNR 過濾器 (141)灰階方框像素元 (14 2)彩度方框像素元 (151)步驟一 (1 5 2)步驟二 (1 5 3)步驟三 (1 5 4)步驟四 Ο (15 5)步驛五 (1 5 6)步驟六 14Motion picture input device boundary detection module interpolation filter image intensity mask matrix high resolution interpolation module image output device (10 0) image frame (101) brightness intensity (10 2) two-dimensional matrix (1 1〇 JBD group (1 1 1 ) JBD high-pass filter (112) image frame correlator (1 3 0 ) JBD interpolation operation module 201044877 (13 2) interpolated pixel points (131) interpolated reference pixels (121) Original pixel point (12 2) New pixel point (12 0) 2D pixel frame (12 4) Horizontal ring type column (1 2 3) Vertical ring type array 〇 (1 4 ◦ ) JBD—SNR Filter (141) Grayscale Box Pixel Element (14 2) Chroma Box Pixel Element (151) Step 1 (1 5 2) Step 2 (1 5 3) Step 3 (1 5 4) Step 4 (15 5) Step 5 (1 5 6) Step 6 14