1254216 九、發明說明： 【發明所屬之技術領域】 /本發明係有關於-種全晝面運動向量估測的方法，應用於 處理系統’特別是-種先進行篩選區塊來進行全 向置估測的方法。 一 【先前技術】 由於科3的日益進步，人的生活水準也跟著提高，對於各 種夕媒體的資訊取得也迫切的增加，對於晝質和速率的要求也 ，來越高，-般最常見到的問題峰拍攝時手部晃動，導致 ^像晃動不清晰。故，目前市面上的影像_產品多半會且有 數位防手振（Digital Image StabiliStiQn，町 力此，其最主要的基本概念係為利用全晝面運動向量 Motion Vector，以下簡稱GMV)估測，來推斷出拍 =晃動的方向並加以補償回復，以能達到更高影像壓縮率且 二^失真的需求。而在運動向量估測領域中，以區塊比對估 niock-basedMatching)最多人使用，此方法係比較兩張 钟旦面之中，兩個不同區塊之間的絕對差值總和（Sum ；以下簡稱為SAD值），得到最小SAD Z區塊位置和原先區塊位置的差值，即是估測的運動向量。 為此方法複雜度較低且較具規律性，適合設計成硬體。 為_了求得GMV，常見作法是把區塊均勻散佈在整張晝面當 :叙H一個區塊都經過區塊比對的運算後，得到對應的局部 (L〇Cal Μ〇ΐί〇η VeCt〇r，以下簡稱 LMV)。之後在經 ϋ平ί的方式’以取得gmv。然而，有些區塊因為晝面内 ^ B而得到不準確的LMV。因此，為了不影響GMV的準確度， ^二不準的LMV要被篩掉，不要拿進來一起做加權平均的運 但是，那些會算出不準LMV的區塊如果能在做Ik Cmilg之前就可以先被發現且排除在外的話，勢必會節省 1254216 許多運算的複雜度。 刖案如美國專利公告第4, 984, 074、5, 748, 231、 5, 563, 652號專利，是先所有的區塊都做完區塊比對之後，再 從區塊比對所得到的一些統計值來做為篩選區塊的依據，他們 ^別汉什了二個不同條件來刪去不適合的區塊，但其效果並不 '’篩選的動作只利用了區塊比對所產生的統計 =亚 >又有利用影像本身㈣訊來做判斷，容易影響判斷的準 何’因為都是採用先計算所有區塊計算的方式， :、、、'後再將不要的區塊刪除，故運算上負擔相當大。 me 面’因為影像要進行儲存前多須經過壓縮，又經過 影像是未來發展駐流，因此如何把兩種技術 ϋ仕、目丨當務之急。由於DIS和影像壓縮都需經過區塊 的计异步驟，前案如美國專利公告第6, 628, 711號專 縮，而嗜计异後所求得㈤直接拿來利用於影像壓 ^法，^，需要再次計算LMV的運算。其所採用的 LM拿轉重疊區域最大者，將前面影像之 子測所求得的最佳運動向量，故影像壓_品質難以提^。 【發明内容】 運動本發明的主要目的在於提供—種全書面 法，能夠以較少舰 測的i達上述目白勺’本發明所揭露之全晝面運動向量估 不用花ΐ餘的3$具+有足夠對比特徵的位置來放置區塊，而 二餘的運异在私沒有什麼内容特徵的區塊。 j，此：二=二進=器的而；分計 有邊角、高對比等具高特徵的區出量具 I254216 而後藉由各區塊的局部運動向量來求算出全晝面運動向量。 另一方面，也可以配合相機系統固有的自動對焦系統與自 動曝光系統，利用其統計值來作為選擇區塊擺放位置的依據， 更能進一步減低計算的負擔。 為使對本發明的目的、構造、特徵、及其功能有進一步的 瞭解，茲配合實施例詳細說明如下。 【實施方式】 根據本發明所揭露之全晝面運動向量估測的方法，有梦於 I知技術上需要針對整張待測影像1〇佈滿區塊並一一加=計 异分，，請參閱「第1圖」，然而，就其中某些區塊來說，壁 t第-，101來說’如「第2圖」所示’其影像内容具^ f個重稷週期性的圖案，觀察其作區塊比對估測後所 SAD值’為SAD值（縱軸）相對於位置（平面的兩轴）的&化， 可以明顯看出其SAD值的變化也係呈現週期性，因此曾与 像運動向斜，難讀測其正確的勒向量。另外，链 特徵、或是顏色分佈均勻或接近等區塊，也容^估^ 運動向量的狀況，故，在推算全晝面運動向量時 ί t的t，拿辭絲贿㈣塊麵μ ί體ί 的運動向1。然而這樣的方法必須先經過區塊比 ^理’制用得觸騎絲進行刪除效果不佳區塊^ 雜職爾上_區塊比對估 而耗費非书夕的運异）。故本發明朝向此 舁影像之區塊的運動向量之前，紐篩選出具 $ ’譬如為第二區塊1〇2，請同時參閱「第 乂右二 f針對此塊純計算，來大幅降低所需要計算的^』 #牛的全晝面運躺量估_方法，其主 要步驟抓“茶閱「弟4圖」，首先接收第一影像與第二^ 1254216 C步驟401)，此虚的楚_ 譬如為單-照片等）中，相U第二影像可為靜態影像（可 振補償的依據，也可以是動H兩張影像，用來作為防手 拍攝連續影料《的彡片等）所 特徵内容的位置設置區播Γ者直接於第-衫像中選取其高 每-區塊的局部運動；量(‘驟行求取所選擇 動向量來求算全晝面運動』量,4二3)’而可猎由這些局部運 續可藉由齡晝面物向量^^ 40ϋ P應用上，後 405)，而可輪出對弟一影像進行補償（步驟 内容位置的方下針對影像中選取高特徵 首先’如「第5A圖所 、 出複數個等分的區塊（步驟5 用H M選取 計算影像内容當中邊〇緣6的強0per咖等 際作法兔趑― 、’強度等，來師述出適合的區塊。其實 Γ —域波11做迴旋積分(贿QMi〇n) 運开（步驟502)，濾波器的形式如下面所示： η； -1 -1 〇 「 -1 0 1 ^ 1 1 -1 0 1 --1 0 Γ --1 -1 —Γ 一0 1 1一 _-1 -1 0 一1 0 1 _-1 0 1 0 0 0 _! 1 1 波器’第一、二個滤波器係分別為45度角和135 二第三、四個濾、波器係分別為0度角和90声角 因生活的場景大多以垂直和水平的邊線^ 塊是由同日寺包含垂直和水平方向的邊線所構成， 哭沾二，i比僅由水平或垂直的邊線所構成的區塊，對於濟波 疋由早方向的邊緣（uni-directional edge)所組成。 i-角所構成的影像，在經過區塊比對估測的運算後，得到 I254216 的LMV正確度會比較高。因此，藉由上述濾波器可將具有邊 的區塊加以篩選，例如显殖使用第一個和第二個濾波器， 座使用第三個和第四個，選取對兩個不同濾波器的響應值 大的區塊，當作適合使用的區塊，因為此類比較有可能是由^ 角構成。使用兩個濾波器會比只使用一個效果來的更好。1254216 IX. Description of the invention: [Technical field to which the invention pertains] / The present invention relates to a method for estimating the motion vector of a full facet, which is applied to a processing system, in particular, a screening block for omnidirectional positioning Estimated method. A [previous technique] Due to the progress of the department 3, the standard of living of people has also increased. The information acquisition of various eve media has also increased steadily. The higher the requirements for enamel and speed, the more common. The problem peaks when the hand is shaking, causing the image to be unclear. Therefore, most of the current image_products on the market have digital anti-vibration (Digital Image StabiliStiQn, the main basic concept is the use of the full face motion vector Motion Vector, hereinafter referred to as GMV) estimation, To infer the direction of the beat = sway and compensate for the recovery, in order to achieve higher image compression rate and the need for distortion. In the field of motion vector estimation, the most use of the block-based estimation (nock-based Matching) is compared. This method compares the sum of the absolute differences between two different blocks in the two clock faces (Sum; For the SAD value), the difference between the minimum SAD Z block position and the original block position is obtained, which is the estimated motion vector. For this reason, the method is low in complexity and regular, and is suitable for designing into a hard body. For the GM to obtain the GMV, the common practice is to spread the block evenly over the entire surface. When a block of Syria H passes through the block comparison operation, the corresponding part is obtained (L〇Cal Μ〇ΐί〇η VeCt〇r, hereinafter referred to as LMV). Then in the way of 经 ί 以 to get gmv. However, some blocks get inaccurate LMV because of the inside of the face. Therefore, in order not to affect the accuracy of the GMV, the two LMVs that are not allowed to be screened out, do not take them together to do the weighted average. However, those blocks that can calculate the LMV can be used before Ik Cmilg. If it is first discovered and excluded, it will inevitably save 1254216 many operations complexity. For example, U.S. Patent Nos. 4,984, 074, 5, 748, 231, 5, 563, 652, after all the blocks have been tiled, and then obtained from the block comparison. Some statistical values are used as the basis for the screening block. They have two different conditions to delete the unsuitable blocks, but the effect is not ''screening action only uses the block comparison to produce The statistics = sub-> use the image itself (four) to make judgments, which easily affects the quasi-judgment of judgments. 'Because they all calculate the calculation of all blocks first, :,,, and then delete the blocks that are not needed. Therefore, the computational burden is quite large. Me face] Because the image needs to be compressed before being stored, and the image is the future development of the image, so how to put the two technologies into consideration and witness the urgent task. Since both DIS and image compression need to go through the block calculation steps, the previous case is as disclosed in US Patent Publication No. 6, 628, 711, and the ambiguity is obtained after the difference (5) is directly used for the image compression method. ^, you need to calculate the LMV operation again. The LM used to take the largest overlap area, the best motion vector obtained by the previous image, so the image pressure _ quality is difficult to improve. SUMMARY OF THE INVENTION The main purpose of the present invention is to provide a full-text method capable of estimating the full-surface motion vector disclosed by the present invention with less ship test. + There is enough position to compare features to place the block, and the second difference is in the private block with no content features. j, this: two = two into the device; sub-meters with corners, high contrast and other features with high features I254216 and then calculate the full face motion vector by the local motion vector of each block. On the other hand, it can also be used in conjunction with the autofocus system and automatic exposure system inherent in the camera system, and its statistical value can be used as the basis for selecting the placement position of the block, which further reduces the computational burden. In order to further understand the objects, structures, features, and functions of the present invention, the embodiments are described in detail below. [Embodiment] According to the method for estimating the full face motion vector disclosed by the present invention, there is a dream that the technology needs to fill the block for the entire image to be tested and add one by one. Please refer to "Figure 1". However, for some of the blocks, the wall t-, 101 is 'as shown in Figure 2'. The image content has a repeating pattern of Observing the SAD value after the block comparison is estimated as the SAD value (vertical axis) relative to the position (the two axes of the plane), it can be clearly seen that the change of the SAD value is also periodic. Therefore, it has been oblique with the image motion, and it is difficult to read its correct Le vector. In addition, the chain characteristics, or the color distribution is even or close to the block, and it is also possible to estimate the condition of the motion vector. Therefore, when estimating the full kneading motion vector, t t, take the bribe (four) block surface μ ί The movement of the body ί is toward 1. However, such a method must first use the block to compare the use of the tactile wire to remove the poorly performing block ^ miscellaneous erg _ block comparison and estimate the cost of non-book eve. Therefore, before the present invention moves toward the motion vector of the block of the image, the New Zealand filter produces $', for example, the second block 1〇2, please also refer to "the second right f for the block pure calculation, to greatly reduce the need Calculated ^ 』 #牛's full face 运 evaluation _ method, the main steps to grasp the "tea reading " brother 4 map", first receive the first image and the second ^ 1254216 C step 401), this virtual Chu _ For example, in a single-photo, etc., the second image of the phase U can be a still image (the basis of the vibration compensation, or two images of the moving H, used as a hand-off film for continuous filming, etc.) The location of the feature content sets the regional broadcaster to select the local motion of the high per-block directly from the first-shirt image; the amount (the sudden selection of the selected motion vector to calculate the total kneading motion), 4 3) 'And can be hunt by these partial runs can be compensated by the age of the face vector ^ ^ 40 ϋ P application, then 405), and can turn out the image of the brother to compensate (step content position in the image for the image Select the high feature first, such as the block in Figure 5A, which is divided into several equal parts (step 5 is selected by HM) Calculate the content of the image in the edge of the edge of the strong 0per coffee, such as the rabbit 趑 、, 'strength, etc., to explain the appropriate block. In fact Γ 域 域 域 domain domain 11 do swing points (Brit QMi〇n) Step 502), the form of the filter is as follows: η; -1 -1 〇 " -1 0 1 ^ 1 1 -1 0 1 --1 0 Γ --1 -1 —Γ 一0 1 1一_ -1 -1 0 1 1 0 1 _-1 0 1 0 0 0 _! 1 1 Waveper 'The first and second filter systems are 45 degree angle and 135 two third, four filter, wave system The 0 degree angle and the 90 sound angle are mostly vertical and horizontal edges. The block is composed of the same day and the horizontal direction of the same day. The cry is two, i is only horizontal or vertical. The block formed is composed of the uni-directional edge of the zibo 。. The image formed by the i-angle is obtained by the block comparison estimation operation, and the LMV correctness of the I254216 is obtained. It is relatively high. Therefore, the blocks with edges can be filtered by the above filter, for example, the first and second filters are used for the display, and the third and fourth are used. Take a block with a large response value for two different filters as a suitable block, because such comparisons may consist of ^ corners. Using two filters is better than using only one effect. .
接著將算出來的值經過絕對值計算（步驟5〇3)，最後p 到影像對不同方向的濾波器所產生的響應值(端視其由 J 波器來計算）；然而，前面所算出的響應值必須經過正規= (Normalization)的處理（步驟504)，以去除亮度的 實際應用於影像10，利用過濾器加以篩選後，可以 如「第6圖」其中8個具有明顯特徵的區塊，相較於方 (第1圖），其所需要計算28個區塊來比較，計算的負^ 顯減輕許多。 貝擔上明 #西另一貫施例’凊參閱「第5β圖」’同樣先將第二影像先定 義遥取出複數個等分的區塊（步驟咖） 導 曝光的動作，其分別是#由自動對自，焦、 System’以下簡稱AFS)鱼自叙22：；、、糸統（Aut0 FocusThen the calculated value is calculated by the absolute value (step 5〇3), and finally p is the response value generated by the image to the filter in different directions (the end is calculated by the J waver); however, the previously calculated The response value must be subjected to normal = (normalization) processing (step 504), to remove the actual brightness applied to the image 10, and after filtering by the filter, as shown in "Fig. 6", among the eight blocks having obvious features, Compared to the square (Fig. 1), it is necessary to calculate 28 blocks for comparison, and the calculated negative is significantly reduced. Beaconshang Ming#West another example of ''see the "5th figure"", the second image is first defined to remove the plurality of equally divided blocks (step coffee) to expose the exposure, which are respectively # Automatic pairing, coke, System' hereinafter referred to as AFS) fish self-report 22:;,, SiS (Aut0 Focus
System ^ AES) 高頻成分值來作對焦，而影像I中^ :依據影像内容的 高頻成分值必定彳…故和邊緣的區塊，其 影像亮度高低的依據，因此加總值來當作評估 則建所叶异得到的AFS值除以 10 1254216 AES值，即達到正規化（N〇rmalizat 然後可篩選出所需要的區塊（步驟604)。果（力驟6 ) k取元區塊後，由第一影像中的區盘一 :〜置的，_比對估測，而可求取出局;運;System ^ AES) The high-frequency component value is used for focusing, and the image I is ^: according to the high-frequency component value of the image content, the image of the edge of the image must be based on the height of the image, so the total value is taken as The evaluation then divides the AFS value obtained by the leaf division by the value of 10 1254216 AES, that is, normalization is achieved (N〇rmalizat can then screen out the required block (step 604). Fruit (force 6) k after the element block , by the first image in the disc one: ~ set, _ comparison estimation, and can be taken out of the bureau;
Motion Vector ； LMV) ( 403 ). Z r 7A " 影像10中之區塊内所绔示的古A >考弟5、7A圖」’ 辦播所;登少W士 r Γ日/、、向P為局部運動向量’而後在 根據所k取之所有區塊的局部 向量（Global Motion Vector · =?止木求取出全晝面運動 LMV 4〇4) ^Motion Vector ; LMV) ( 403 ). Z r 7A " Ancient A > Cousin 5, 7A Picture" in the block of Image 10" Broadcasting Station; , to P is the local motion vector ' and then in the local vector of all the blocks according to the k (Global Motion Vector · =? stop wood to extract the full face motion LMV 4〇4) ^
圖」與「第7B C ?第?B疋圖加=均t方式。請比對「第7A 區塊計算的方式，故其總」全區域劃分 量，且好區油為卿在^=8個區塊的局部運動向 7A圖」為利用好的區塊去除掉。而「第 塊的局部運動向量，由鬼的方式，、僅需要計算8個區 所以經過區塊崎估測得^ = 6她過區塊的篩選步驟’ 做進-步_選，㈣省、=，準確率會比較高，不必再 體計算上可以省去相當多外2G個區塊的運算量，故整 AES值的導入，更進—牛、^十异負擔，且配合利用AFS值與 v，去使用的資源。 另一方面，影像擷取梦 縮儲存，來節省儲存的办^ :存影像時，多半會將其加以壓 都需要做運動向量估‘ ;°而影像壓縮與DIS的共同特徵是 計算量。運用本發明前述所斗二本發明將兩者加以結合來節省 子（predictor)，請參閱「°1异出的全晝面運動向量作為預測 個區塊8H，利用前述方法弟8圖」’原始影像81包含有複數 四個區塊811來計算求取出可，用較少的區塊，以本圖為例， 面補償後得到待壓縮影像R其全畫面運動向量83，之後作晝 行處理時，每一區塊821 Y，故，當需要對待壓縮影像82進 預測子，故朝向全書面運用此一全晝面運動向量83作為 一由運動向量83的方向優先進行尋找是^ 1254216Figure" and "7B C?第B疋图加=均特的方式。 Please compare the "7A block calculation method, so its total" is divided into the whole area, and the good area oil is Qing in ^=8 The local motion of the blocks is removed from the 7A map for the good blocks. And "the local motion vector of the first block, by the ghost way, only need to calculate 8 regions, so after the block estimation estimate ^ = 6 her screening process for the block" to do - step _ election, (four) province, =, the accuracy rate will be relatively high, and it is not necessary to recalculate the calculation amount of a considerable number of 2G blocks. Therefore, the import of the entire AES value is more intensive, and the AFS value is used together with v, to use the resources. On the other hand, the image captures the dream storage, to save the storage of the device ^: When storing the image, most of them will be pressed to do the motion vector estimation '; ° and image compression and DIS The common feature is the amount of calculation. By using the two inventions of the present invention to combine the two to save the predictor, please refer to "the full face motion vector of °1 is used as the prediction block 8H, using the aforementioned method. 8) "The original image 81 contains a plurality of blocks 811 for calculation and extraction, and fewer blocks are used. Taking this picture as an example, the face image is compensated to obtain a full picture motion vector 83 of the image to be compressed R. After the processing, each block is 821 Y, so when needed The compressed image 82 is subjected to the predictor, so that the full face motion vector 83 is applied to the whole writing as a direction by which the motion vector 83 is preferentially searched for ^ 1254216
【圖式簡單說明】 ^ 1圖係為習知區塊比對估耻之區塊擺放的示意圖。 f 2圖係為習知區塊中具週期性圖樣之示意圖。 f3圖係為習知區塊中具高特徵值區塊之示意圖。 圖。第4圖係為本發明全晝面運動向量估測之步驟流程示〕 二5A 5B ®係為本發明區塊選擇之步驟流程示意圖。 第6圖係為本發明影像中選擇區塊之示意圖。 >第7A、7β關林發_習知技術之比較圖。 第8圖係為應用本發_縮影像之示意圖。 ίο主要元件符號說明】 101 影像 1〇2 81 811 82 第一區塊 第二區塊 原始影像 區塊 12 1254216 821 83 區塊 全晝面運動向量[Simple description of the figure] ^ 1 is a schematic diagram of the block placement of the conventional block. The f 2 diagram is a schematic diagram of a periodic pattern in a conventional block. The f3 diagram is a schematic diagram of a block with high eigenvalues in a conventional block. Figure. Figure 4 is a flow chart showing the steps of the full-surface motion vector estimation of the present invention. The second 5A 5B ® is a schematic flow chart of the steps for selecting the block of the present invention. Figure 6 is a schematic diagram of selected blocks in the image of the present invention. > 7A, 7β Guan Linfa _ comparison chart of conventional technology. Figure 8 is a schematic diagram of the application of the present invention. ίοMain component symbol description] 101 image 1〇2 81 811 82 First block Second block Original image Block 12 1254216 821 83 Block Full face motion vector
1313