200929059 九、發明說明： 【發明所屬之技術領域】 本發明係有關一種相片管理的系統與方法，特別是關 於種利用影像辨識進行相片分類的系統與方法。 【先前技術】 數位相機的技術近年來發展的非常 Ο ❹ 提升到幾乎以全取代傳統光學相機〜，: 相片斷增大。然而’目前的數位相機可儲存的 受到儲存容量的：f丨，讓使用者可以盡情拍攝，不需擔心 上的困難a 卻也增加了使用者在相片分類整理 攝時間八翻善此一問題，目前的數位相機可以根據拍 能，以相 或者藉由相機内建的衛星定位系統(GPS)功 言仍不夠方^拍攝地點做分類，但這樣的分類對使用者而 片頃Ϊί二ί位相機的使用者習慣於將每次拍攝完成的相 大的相片:高的儲存媒體’例如硬碟，在累積數量較 變得困難，/類整理的工作變得困難，搜尋特定相片也 片。雖缺J在往需要花費很長的時間瀏覽數量極多的相 是仍然：法多圖像管理軟體可以簡化部分作業，但 龐大數=的主題或人物為相片分類。對於擁有 是辛苦且費日^的管理者而言，為相片進行分類的工作也 另〜方面，雖然影像辨識技術已經有报好的發展，但 200929059 ::關r尚未具有辨識特定 作也常:使在一般的電腦系統上’辨識特定人物影像的： 化費高度的系統資源及很長的時間。對於相片分 碑决方ί的工作來說，先進的影像辨識技術並不是合適的 Ο 此’一種簡單、便宜的相片分類系統及方法，乃為 c、内容】 在於提出-種利用影像辨識進行 ， 4未發明的目的之 分_的系統。 〜的之一，在於提出1利用影像辨識進行 分細來發明的目的之一，在於接屮一弘 的程式竭的電腦可讀取媒體。儲存有供進行相片 迷隹、本發明…種自動化相片管理系 相片：; ⑽體’第-記憶體用來提供許多需=處理器 相〜：多相片進行搜尋及辨識，找根據該程式 =較高的相片歸類為同一組。找出具有某些影像特徵 相片2本發明，i自動化相片 4覆出現的影像特徵 4包括辨識許多 及根據該影像特徵將4 200929059 多相片分類。 化-用來判斷相似度的影 色或=太例如相片中的人臉、服飾St影像區塊的顏 一處理器讀取，简行—程序，，有程式竭，供 中重覆出現的影像特徵，據=辨識許多相片 片分類。 影像特徵將該許多相 Ο 【實施方式】 園1係根據本發明的自動 包括處理器12連接，二 片g理系統的實施例， Z逆楼5己憶體14及16。第— 揮發性記憶體，例如 》憶體14是非 π如硬碟或快閃記憶卡，用來 行分類的相片。第-# &μ μ 用來叔供需要進 己憶體16是唯讀記憶體，用夾儲在 程式碼。另外，如同—船㈣用來儲存 股的系統’還有一個第：r#,陪磁 是隨機存取記憶體，用㈣❹“第— 6&_ 18 用於辅助處理器12運算所需。處理 〇 器12根據第二記憶體16中的程式碼對第-記憶體14中 的相片進行搜尋及_，找出具有某些影像特徵相似度較 高的相片歸類為同一組。 在進行影像辨識以前，處理器12可以對相片進行初 分類，例如依據相>1的拍攝日期或拍攝地點先將相片區分 為相同拍攝日期或拍攝地點類別。 在進行影像辨識時，處理H 12也可以先依據相片中 出現的人臉數目進行初分類，例#人臉數目相同或低於某 一數量的相片歸為同一類。 7 200929059 在利用影像辨識為相片 影像辨識，只利用某此影像二:步为類時’不必作精細的 相片歸為同-組。這:=徵的相似度來判斷是否要將 狀，例如服飾或配件特徵包括影像區塊的顏色或形 飾特===:^，，有相同的服 相似度辨識。 、，再對同組中的相片進行人臉的 在200929059 IX. Description of the Invention: [Technical Field] The present invention relates to a system and method for photo management, and more particularly to a system and method for photo classification using image recognition. [Prior Art] The technology of digital cameras has developed very 近年来 近年来 in recent years. It has been upgraded to almost completely replace traditional optical cameras~,: Phase segments have increased. However, 'the current digital camera can store the storage capacity: f丨, so that users can enjoy shooting, no need to worry about the difficulty a, but it also increases the user's photo sorting time eight good. The current digital camera can be classified according to the beat energy, or by the camera's built-in satellite positioning system (GPS), but it is still not enough to classify the shooting location, but such classification is for the user and the film is 片ίίί camera Users are accustomed to the large photos that are completed each time: high storage media such as hard disks, which become difficult to accumulate, work hard to sort, and search for specific photos. Despite the fact that it takes a long time to view a large number of phases, it is still possible: Fado image management software can simplify some operations, but the huge number = theme or character is a photo classification. For managers who are hard-working and cost-effective, the work of classifying photos is also another aspect. Although the image recognition technology has been reported to be well developed, 200929059::n is not yet recognized for specific work: Enables the identification of specific person images on a typical computer system: a high level of system resources and a long time. For the work of photo-sharing, the advanced image recognition technology is not suitable. This is a simple and inexpensive photo classification system and method, which is c, content. It is proposed to use image recognition. 4 The system of the uninvented purpose. One of the ~ is to propose that one of the purposes of using image recognition for thinning is to access the computer-readable media. Stored for photo confusion, the invention...Automatic photo management system photos:; (10) Body 'first-memory is used to provide many needs = processor phase ~: multiple photos for search and identification, find according to the program = compare High photos are grouped into the same group. Finding a photo with certain image features 2 The present invention, the image feature 4 that appears in the overlay photo 4 includes identifying a plurality of images and classifying 4 200929059 multiple photos according to the image features. - used to determine the similarity of the color or = too, for example, the face in the photo, the clothing of the St image block of the processor - read, simple - program, there are exhaustive, for the medium and repeated images Features, according to = identify many photo film categories. The image features are many of the same. [Embodiment] The garden 1 according to the present invention includes an embodiment in which the processor 12 is connected, a two-chip system, and a Z-reverse building. The first - volatile memory, such as "Recall" 14 is a non-π such as a hard disk or a flash memory card, used to sort photos. The ## &μ μ is used for the uncle to be used. The memory 16 is a read-only memory, which is stored in the code. In addition, as in the case of the ship (four) used to store stocks, there is also a first: r#, the accompanying magnet is a random access memory, and (4) ❹ "6-6" is used for the auxiliary processor 12 operation. The device 12 searches for the photos in the first memory 14 according to the code in the second memory 16, and finds that the photos with higher image similarity are classified into the same group. In the past, the processor 12 can initially classify the photos, for example, according to the shooting date or the shooting location of the phase > 1 to first distinguish the photos into the same shooting date or shooting location category. When performing image recognition, the processing H 12 can also be based on The number of faces appearing in the photo is initially classified. For example, the number of faces with the same number of faces or less than a certain number is classified into the same category. 7 200929059 In the use of image recognition for photo image recognition, only one image is used: When you are class, you don't have to make a detailed photo into the same-group. This: = the similarity of the sign to determine whether you want to change the shape, such as clothing or accessories features including the color of the image block or the shape of the special ===:^, Have the same Similarity identification. ,, again for the same group of photos in the human face
胸Μ實例中’如果同組中的相片人臉的相似度低， 具^配件與衣物的比對比重，重新利用服飾作筛選，將 2相同的服飾特徵的相片分為同一組，再對同組中的相 遵行人臉的相似度辨識。 户，處理器12也可以比較不同組的相片中人臉的相似 如果不同組別中的人臉具有高相似度則判斷不同組 、相片屬於同一人，進而將這些組別連結在一起。 圖1的實施例可以在數位相機或其他裝置 個人 電鵰）上實現。 圖2係根據本發明的自動化相片管理方法的第一實施 例的流程圖，在本實施例中，係利用數位相機在沒有拍照 的空檔時間，對儲存於相機内的相片進行分類處理。首先 進行步驟101 ’將相片依據拍攝日期做分類，並在步驟1〇2 中篩選出單人的獨照。由於在同一天中，同一人幾乎都會 穿著相同的服飾，因此再進行步驟103，將同一日期的相 中具' 有相同服飾特徵的單一人物相片設定為同一組，以 提為同一組相片中為同一人的機率。接著，進行步驟104， 8 200929059 挑選出同一組相片中人臉特徵最清楚的數張相片做人臉 辨識，人臉辨識技術係藉由比對鼻子、嘴巴、眼睛、臉型 等進行相似度比對而加以辨識，這些已是習知技術，在此 不多加贅述。在步驟105中，若同一組相片中的人臉相似 度高於門檻值’便可以判斷該組相片中的人物為同一人， 若否’則表示該組相片中的人物是穿著相似衣物的不同 人’因此進行撞衫處理裎序1〇6，在步驟107中調整配件 與服飾的比對比重，例如提高眼鏡、手錶、項鍊等飾品配 ® 件的比對比重’藉以將同一類中的相片重新做分組’再進 行人臉的相似度辨識，若重新分組後，人臉相似度仍低於 門檻值，如此重覆達一定次數後便進入步驟11〇，不再對 這些相片做處理。回到步驟105，判斷出同組相片中的人 物為同一人後，進行步驟108，從不同組的相片中選取一 張人臉辨識度最尚的相片，互相做相似度辨識，在步驟1〇9 中，若相似度低於門檻值，表示各組相片中的人物分別為 〇 不同人’維持原本的分組及連結，不做處理 ;若不同組相 片中的人臉經辨識後，相似度高於門檻值，表示這兩組相 片中的人物是同一人，進行步驟lu，將這兩組相片判斷 為同一人的相片，並互相連結’以供使用者選取。較佳者， 可以將每組相片中人像特徵最清楚的人臉部份做為圖 示，儲存在相機的内建記憶體中，供使用者點選，便於未 來搜尋相片。 圖3係圖2之實施例的應用例示意圖，數位相機2〇 上有多個按鍵32、34、36、38及方向鍵4〇，用來選取或 200929059 =存機二:::或模式，、播〜 後，編:=:::::::經广之“: 為選擇圖示，例如第一組相片的代表人像=人:部份做 第二組相片的代表人像為第二人像 ：24 ’ =三人像第四組相片的代表:像 田使用者在顯示螢幕22上選擇第二人像26時 €)In the chest example, 'if the similarity of the photo face in the same group is low, the proportion of the matching of the accessories and the clothing is re-used, and the photos of the same clothing features are divided into the same group, and then The similarity in the same group is recognized by the similarity of the face. The processor 12 can also compare the similarities of faces in different sets of photos. If the faces in different groups have high similarity, it is judged that different groups and photos belong to the same person, and then these groups are linked together. The embodiment of Figure 1 can be implemented on a digital camera or other device. 2 is a flow chart of a first embodiment of an automated photo management method according to the present invention. In the present embodiment, a digital camera is used to classify photos stored in the camera during a blank time without photographing. First, proceed to step 101 ’ to sort the photos according to the shooting date, and filter out the single photos in step 1〇2. Since the same person wears the same costume almost all the same day, step 103 is performed to set a single person photo with the same costume feature in the same date as the same group, so as to be the same group of photos. The odds of the same person. Then, in steps 104, 8 200929059, a plurality of photos with the most clear facial features in the same set of photos are selected for face recognition, and the face recognition technology is compared by comparing the nose, the mouth, the eyes, the face, and the like. Identification, these are already well-known techniques, and will not be repeated here. In step 105, if the face similarity in the same set of photos is higher than the threshold value, it can be judged that the characters in the set of photos are the same person, and if not, the characters in the set of photos are different in wearing similar clothes. The person's therefore handles the shirt order 1〇6, and in step 107, adjusts the specific gravity of the accessories and the clothing, for example, increases the proportion of the accessories of the glasses, watches, necklaces, etc. to reproduce the photos in the same category. Do grouping' and then perform face similarity identification. If the grouping is similar, the face similarity is still lower than the threshold. After repeating the number of times, the process proceeds to step 11 and the photos are no longer processed. Going back to step 105, after determining that the characters in the same group of photos are the same person, proceeding to step 108, selecting a photo with the most facial recognition degree from the photos of different groups, and performing similarity recognition with each other, in step 1 9 If the similarity is lower than the threshold, it means that the characters in each group of photos are different people's original grouping and linking, and no processing is done; if the faces in different groups of photos are recognized, the similarity is high. The value of the threshold indicates that the characters in the two sets of photos are the same person. Steps lu are used to determine the two sets of photos as photos of the same person and are linked to each other for selection by the user. Preferably, the face portion of each group of photos with the most clear portrait features is displayed as an image stored in the camera's built-in memory for the user to click to facilitate searching for photos in the future. 3 is a schematic diagram of an application example of the embodiment of FIG. 2. The digital camera 2 has a plurality of buttons 32, 34, 36, 38 and a direction key 4 〇 for selecting or 200929059=storage two::: or mode. After the broadcast ~, edit: =::::::: by Guangzhi": For the selection of icons, for example, the representative image of the first group of photos = person: part of the representative image of the second group of photos is the second portrait :24 ' = Three people like the representative of the fourth group of photos: like the field user selects the second portrait 26 on the display screen 22)
圖4係根據本發明的自動化相片管理方法的第二實施 例的流程圖’係配合使用者在出遊時常會出現情侣合照的 習慣所設計的-種管理雙人合照的方法。和圖2—樣，先 進行步驟5(H，將相片依據拍攝日期分類，再經步驟， 從同類相片中筛選出出現兩個人的相片，再對這些相片進 行步驟503，依據服飾特徵分組，而後對同一組中的相片 進行人臉的相似度辨識504，在步驟505中，若相似度不 咼於門檻值’和圖2之實施例相同地，進入撞衫處理程序 506,進行重新分組507或者不做處理51〇;若相似度高於 門檻值，便可判斷該組相片中的人物為同一對人物，再進 行步驟508 ’從不同組的相片中各取出一張人臉辨識度最 高的相片，互相做相似度辨識，在步驟5〇9中，若相似度 低於門權值，表示各組相片中的人物不是同一對人物’維 持原本的分組及連結，不做處理51〇，若不同組相片中的 人臉經辨識後，相似度高於門檻值，表示這兩組相片中的 200929059 人物是同一對，進行步驟511，將這兩組相片判斷為同一 對人物的相片，並互相連結以供使用者選取。 以上的實施例以數位相機為例，只是方便解說，本發 明並不侷限於數位相機。 【圖式簡單說明】 圖1係根據本發明的自動化相片管理系統的實施例； 圖2係根據本發明的自動化相片管理方法的第一實施 Ο 例的流程圖； 圖3係圖2之實施例的應用例的示意圖；以及 圖4係根據本發明的自動化相片管理方法的第二實施 例的流程圖。 【主要元件符號說明】 12 14 16 18 101 102 103 〇 處理器 第一記憶體 第二記憶體 第三記憶體 將相片依據拍攝日期分類 自同類相片中篩選出單人獨照 將同類相片中具有相同服飾特徵的人物相片 設定為同一組 104 對同組相片進行人臉的相似度辨識 105 相似度高於門檻值？ 11 200929059 106 撞衫處理程序 107 配件與衣物比對比重調整，重新分組相片 108 自不同組的相片中取一張人臉辨識度最高的 相片進行相似度辨識 109 相似度高於門檻值？ 110 不做處理 111 判斷不同組相片屬於同一人，並互相連結 20 數位相機 Ο 22 顯示螢幕 24 第一人像 26 第二人像 28 第三人像 30 第四人像 32 按鍵 34 按鍵 36 按鍵 〇 38 按鍵 40 方向鍵 501 將相片依據拍攝日期分類 502 自同類相片中篩選出出現兩個人的相片 • 503 將具有相同服飾特徵的相片歸為同一組 504 對一組中的相片進行人臉的相似度辨識 505 相似度大於門檻值？ 506 撞衫處理程序 12 200929059 507 配件與衣物比對比重調整，重新分組相片 508 自不同組的相片中取一張人臉辨識度最高的 相片做相似度辨識 509 相似度大於門檻值？ 510 不做處理 ' 511 判斷不同組相片屬於同一對人物，並互相連結Fig. 4 is a flow chart showing a second embodiment of the automated photo management method according to the present invention, which is designed to cope with the user's habit of taking a couple's photo when traveling. As shown in Fig. 2, first step 5 (H, classify the photos according to the shooting date, and then, through the steps, select photos of two people from the same photos, and then perform step 503 on the photos, according to the clothing features grouping, Then, the similarity recognition 504 of the face in the same group is performed. In step 505, if the similarity is not equal to the threshold value', and the same as the embodiment of FIG. 2, the collision processing program 506 is entered to perform regrouping 507 or If the similarity is higher than the threshold value, it can be judged that the characters in the group of photos are the same pair of characters, and then step 508 'takes a photo with the highest facial recognition degree from each photo of the different groups. In the step 5〇9, if the similarity is lower than the gate weight, it means that the characters in each group of photos are not the same pair of characters' maintenance of the original grouping and linking, no processing 51〇, if different After the face in the group photo is recognized, the similarity is higher than the threshold value, indicating that the 200929059 characters in the two sets of photos are the same pair, and step 511 is performed to determine the two sets of photos as the same pair of characters. The above embodiments are exemplified by a digital camera, but are conveniently illustrated, and the present invention is not limited to a digital camera. [Simplified Schematic] FIG. 1 is an automated photo management according to the present invention. 2 is a flowchart of a first embodiment of an automated photo management method according to the present invention; FIG. 3 is a schematic diagram of an application example of the embodiment of FIG. 2; and FIG. 4 is an automated photo according to the present invention. Flowchart of the second embodiment of the management method. [Description of main component symbols] 12 14 16 18 101 102 103 〇 processor first memory second memory third memory categorize photos from similar photos according to shooting date Single photo alone sets the photos of people with the same costume characteristics in the same photo as the same group 104. Face similarity recognition for the same group of photos 105 Similarity is higher than the threshold value? 11 200929059 106 Sweatshirt processing program 107 Accessories and clothing Comparing the specific gravity adjustment, regrouping the photo 108 takes a photo with the highest facial recognition from the photos of different groups. Similarity 109 Similarity is higher than the threshold value? 110 No processing 111 Judging different groups of photos belonging to the same person and connecting 20 digital cameras to each other 显示 22 Display screen 24 First portrait 26 Second portrait 28 Third portrait 30 Fourth Portrait 32 Button 34 Button 36 Button 〇38 Button 40 Direction Button 501 Sort photos according to shooting date 502 Filter out photos of two people from similar photos • 503 Group photos with the same costume characteristics into the same group 504 The photo in the face is similarly recognized. 505 Is the similarity greater than the threshold? 506 Sweatshirt Processing Procedures 12 200929059 507 Adjusting the Specific Gravity of Accessories and Clothing, Regrouping Photos 508 Taking a Photo with the Most Facial Recognition from Different Groups of Photos for Similarity Recognition 509 Is the similarity greater than the threshold? 510 does not process ' 511 to judge different groups of photos belong to the same pair of characters, and link to each other
1313