TWI450206B - Automatic image identification method - Google Patents

Description

Automatic image recognition method

The present invention relates to an automatic image recognition method, and more particularly to a method for automatically identifying a face or a similar image by a computer system, thereby identifying a face or a similar image source.

Science and technology always comes from human nature, and people are always the most important subject in scientific research. Therefore, in computer vision, research on face is always a hot topic. Currently available digital cameras, mobile phones and other technology products have additional functions to deal with "people" related information. The most widely known functions are face detection, expression recognition, and so on. Unlike these commercial features, face recognition is still a difficult issue. Why is there such a difference? The main reason is that face recognition is provided by digital cameras or smart phones, and the information needed can be applied to everyone, not to specific people. For example, a human face must be two eyes and a nose. By capturing such information, the system can determine that the image is most likely a human face image. However, "face recognition" is different from the above "identification of face". It is usually necessary to first establish a database and let the system know who the person needs to be identified in order to get a corresponding answer through calculation and comparison. In addition, you need to find the image features that belong to everyone to help the system identify. The two eyes, one nose, and the information that belong to the common features of the face, can't play a role in face recognition. Therefore, in face recognition, how to collect "useful" information and how to find out the "characteristics" of each person from these materials is the most important and problem to be solved.

In the past, many research scholars have proposed different techniques to solve the problem of face recognition. But before in the research of face recognition, there is a very strict assumption that the picture when training the face model needs to be completely clean, meaning that there can be no cover like scarves, glasses, masks, etc. to block the face. . Such strict assumptions are less applicable to real-life situations. Because, in real life applications, building faces The face images collected when identifying the database do not require that the images be shot in such an ideal environment. For example, it is usually impossible to obtain a complete face image due to factors such as lighting conditions, obscuration, or image angle.

Therefore, it is necessary to provide a novel automatic face recognition method for creating a useful face model for identification purposes based on undesired images.

At the same time, there is a need for a novel method for automatic recognition of face images, which can still make correct recognition under the condition that the characteristics of image data are insufficient.

At the same time, there is a need for an automatic image recognition method that can create useful image models for identification based on undesirable images.

At the same time, there is a need for a novel automatic image recognition method, which can still make a correct identification under the condition that the image data is insufficient.

At the same time, there is also a need for an automatic face recognition system or an automatic image recognition system that can perform the above method.

The object of the present invention is to provide a novel method for automatically recognizing facial images, which can still make correct recognition under the condition that the characteristics of the image data are insufficient.

The object of the present invention is also to provide an automatic image recognition method for creating a useful image model for identification based on an undesired image.

The object of the present invention is also to provide a novel image automatic recognition method, which can still make a correct identification under the condition that the image data characteristics are insufficient.

It is also an object of the present invention to provide a face image automatic recognition system for performing the above method.

It is also an object of the present invention to provide an image automatic identification system for performing the above method.

The image automatic identification method and system disclosed by the present invention proposes a solution for training images that have poor lighting conditions, masking, and segment images when establishing an image model for identification. The method according to the present invention mainly comprises: accurately correcting the error of the pixel by low rank decomposition; and obtaining a correct identification image model, such as a face image model, by separating pixels that do not belong to a specific target. Since the method of the present invention does not have the assumption that the image must be clean, the obtained image can be used to establish the identification image model without pre-processing. Therefore, the present invention does not need to discard some images of poor lighting conditions, obscuration or fragments. In other words, the present invention retains all images and becomes useful information. It's completely different from the previous technology that has to be chosen to discard some of the masked images.

In addition, the present invention introduces the concept of structural dissimilarity to more accurately identify features that belong to a particular subject, such as face, eyes, size and shape of the nose, and the like. The present invention can further improve the recognition rate as compared with the low rank decomposition method alone.

The invention provides a novel image recognition model establishing method, in particular, a method for establishing a face image recognition model, which comprises the following steps: inputting a plurality of digital image files belonging to the n category, each image file comprising a representative image. Pixel address and its grayscale data, and the category information of the image; most image files for each category Di(0<i n) extracting its common feature Ai; calculating a common feature Ai of a class and a common feature of a certain number (m) of other categories Aj (0<j m The difference characteristic n(Aj ^T Ai) of n-1, Dj≠Di); and the difference feature Σ(Aj ^T Ai) is added to the common feature Ai of the category to obtain a model image Ai ^{* of} the category.

Among them, the calculation of the Ai ^* can be obtained by an optimization solution. In a preferred embodiment of the invention, the following formula is calculated via a single step to obtain the Ai ^* :

St D _i = A _i + E _i .

In a preferred embodiment of the present invention, the specific number m is n-1, that is, the common feature of the common feature Ai of the class and all other existing categories are calculated for the difference feature.

In the embodiment of the present invention, the method of obtaining the model images A1 ^* to An ^* of each category in the same manner is further included.

The invention also proposes a method for improving the independence of model images by using the same type of training image error images. The erroneous image can be weighted and combined with the model image.

The present invention also discloses a method for determining the degree of approximation of an input image and each type of model image using the obtained model image of each category. The method comprises the steps of: inputting a digital image file to be processed, the image file comprising a pixel address representing an image and gray scale data; comparing the image with the model image obtained by the above method and belonging to most categories Difference, calculating the approximate degree of the image to be processed and the image of each category; and determining the relationship between the image to be processed and the image of each category.

The step of determining the relationship between the image to be processed and each model image includes determining that the image to be processed belongs to the category with the highest degree of similarity. In another example of the present invention, the step of determining the relationship between the image to be processed and each model image includes determining that the image to be processed belongs to a category whose approximation is higher than a threshold. In most instances of the invention, the majority of the digitized image files of the input are facial image files, and the categories from which the face image source is derived. In such an example, the model image created is a genus Image of the face model of most people. However, in other examples of the invention, the model image created is an image representing an item, a venue, or an event.

Hereinafter, an image automatic recognition method of the present invention will be described by way of example, including a method for establishing a model image and a method for identifying a to-be-processed image by using the created model image. It should be noted that the examples of the present invention are directed to exemplifying the method steps of the present invention and the applicable hardware and software components and architectures thereof, enabling experts in the industry to implement them. The detailed description of the present invention is not intended to be exhaustive or to limit the scope of the invention.

The invention provides a novel image automatic recognition technology, in particular, an automatic recognition technology for face images. The invention proposes a method for identifying a to-be-processed image by using a model image, and a high accuracy rate can be obtained.

Fig. 1 is a flow chart showing the method of creating a model image of the present invention. The following is an example of establishing a face model image to illustrate how to create the model image.

As shown in FIG. 1, when the model image is created, first, in step 101, a plurality of digitized image files belonging to the category n (n is a natural number greater than 1) are obtained. In this example, since the purpose is to create a face image comparison or a face model image, the image file obtained at this step is a face image representing a known majority; the so-called n category is Refers to n known characters. The image file obtained in this step usually includes a pixel address representing the image and its grayscale data, and the category information of the image. In other words, in the case of a face image, the obtained image file includes a pixel address representing a face image of a specific person and its grayscale material, and the name or code of the specific person. In face recognition or other applications, it may be necessary to include data representing the color of the pixel and its grayscale. Other materials suitable for image recognition or comparison may also be included in the image file.

In the application of the present invention, the image file is usually a digital image itself obtained by a digital image scanner or a digital camera. However, in other examples of the invention, the digitized image may have been compressed and/or extracted. To reduce the amount of data in the image file or to reduce the dimensions of the image features.

Conventional techniques, when used for face image comparison, have a technique called "Low-rank matrix recovery" to find the same information in the same class. The method is to remove the sparse noise from the original image data D to generate a low rank matrix A for representing the original image. The resulting low rank matrix A preserves the important features of the original image. This method can simplify processing and increase processing speed, and is suitable for various image recognition applications. This method can be used in the present invention as a tool for simplifying processing. Therefore, the image file obtained in step 101 may be an image that has been compressed and/or extracted. If it is the original file of the image, the step of compressing and/or extracting the image may also be added after this step. In application, each image file is preferably adjusted to have the same resolution and size that is multiples, and images of the same orientation. Among the known image processing techniques, there are many image processing software available that can be used to achieve this purpose either manually or automatically. The related art need not be described here.

However, when making facial image recognition or comparison, facial images obtained by capturing different characters usually contain common (related) features such as the location and area of the eyes and nose. For "faces", these image locations or areas are usually similar. Therefore, the generated matrix A may not contain enough distinguishing information. This problem also occurs in other similar image recognition or comparison applications. In order to solve this problem, the present invention provides a method for improving the structural difference between images to emphasize the independence of each image file as a basis for image recognition.

In step 102, the system selects the obtained image file according to its category Di(0<i n) Extract the common feature Ai of the image to which each category belongs. When the common feature is extracted, the pixels with the same coordinates (resolution and size adjustment) in the image file, and the pixels whose gray scale values are close to each other are considered to be common feature points. Other known image feature extraction techniques can also be applied in this step. If the image file is the low rank matrix obtained by the low rank matrix analysis described above, the common features can be directly extracted to obtain the common feature file Ai of each category. The common signature obtained at this time is also a low rank matrix.

Figure 2 shows the result of extracting a common feature from a group of facial images belonging to the same person. In the figure, the upper part is the original image of the group, and the lower part is the obtained common feature image. Of course, Figure 2 is only intended to illustrate an example of the invention and is not intended to limit the invention.

In step 103, a common feature Ai and a specific number (m, m is a natural number) are used to calculate the common feature Aj (0<j). m The difference characteristic of n-1, Dj≠Di) (Aj ^T Ai). Although in the present invention, a useful model image file can be created using a common feature of a few other categories, in the preferred embodiment of the present invention, the common feature image Ai can be combined with other n-1 categories. Compare them one by one to find out the difference between them and all other categories of images. This can improve the accuracy of recognition. In the case where the difference feature Σ (Aj ^T Ai) and the common feature Ai are simultaneously considered in step 104, the feature image Ai ^{* of} the category is obtained by an optimization solution. This completes the creation of a common profile for a category of images. The created common profile can be used as a model image for this category. The completed common feature file has higher independence, can not only represent the common features of the category, but also provide more features, which can improve the correct rate of recognition. In the face recognition application, the common feature file represents a facial feature of a specific person, which is sufficient for identifying whether a pending face image belongs to the specific person.

In a preferred embodiment of the present invention, the model image can be calculated by a single step using only the following formula:

St D _i = A _i + E _i .

At step 105, it is determined whether all of the n types of image files have established a common profile. If no, it will go to step 103. If so, the operation ends in step 106, and the creation of the image common feature files of all categories is completed. The image model file established by the above method has the common feature of the category with improved independence, and the correctness rate can be improved when the image is compared or recognized.

When the model image created by the present invention is used for comparison, a pair of to-be-compared images is compared with the model image that has been archived, and according to the difference value between the two, the category of the pair of images to be compared is determined. In terms of face image comparison, it is the face of the person to whom the image belongs.

Figure 3 shows a flow chart of a comparison method for comparing images using the model images created by the present invention. As shown in FIG. 3, a digital image file to be processed is input in step 301. The image file usually includes a pixel address of an image and its grayscale data. In step 302, the image file is subjected to necessary processing to have a specific resolution, size, and orientation. If necessary, locate the anchor point in the image to increase the speed and accuracy of the alignment. In step 303, a plurality of model image files representing the majority of the images are obtained. Each model image file also includes a pixel address of the image and its grayscale data, and the category information of the image. If the content of the model image is not the pixel address of the image and its grayscale data, but other types, format data, then in step 304 the image to be compared is processed into the same or compatible format. In step 305, the image file to be processed is compared with the feature images of all categories to obtain the difference data. At step 306, a difference value less than a threshold is selected. It is determined at 307 that the digital image to be processed belongs to the selected difference value and the category corresponding to the minimum difference value. This completes the comparison.

Steps 306 and 307 can prevent the selected category difference value from being too large, resulting in a false positive. In fact, after calculating the difference value, the category whose difference value is smaller than the specific threshold value (and the approximation degree is greater than the threshold value) may be selected to determine that the image to be processed belongs to the categories. In addition, it is of course possible to select only the category having the smallest difference value (the highest degree of approximation) and determine that the image to be processed belongs to the category. Face recognition application For example, the judgment result is the face image of the person to whom the image to be processed belongs.

In order to prove the effect of the present invention, the current most credible face database AR Database ( http:///www2.ece.ohio-state.edu/~aleix/ARdatabase.html , Ohio State University, USA) is selected. The facial images of 50 men and 50 women were created and identified by the combination shown in Table 1. Each group of training images has 800 images, and 1,200 images selected from them are used for identification test.

When constructing a model image, the method of the present invention (Ours), low rank matrix analysis (LR), SRC, LLC+SRC, NN, Fisherface method, etc. are used as the basis for comparison. The experimental results are shown in Figures 4, 5, and 6. It is proved that the correct rate of the present invention is superior to the conventional method, and the present invention achieves a relatively high correct rate when the dimension is 75. It is enough to simplify the comparison between the establishment of the model image and the comparison of the images, and improve the speed of image training and comparison.

The above description is intended to disclose the principles, operations, functions and effects of the invention. Instructions The content is only intended to illustrate several preferred examples of the invention. The person skilled in the art can not understand the concept and spirit of the present invention by reading this patent specification, and various changes and extensions are made without departing from the scope of the invention. However, applications such as these are within the scope of the invention.

Fig. 1 is a flow chart showing the method of creating a model image of the present invention.

Figure 2 shows the result of extracting a common feature from a group of facial images belonging to the same person.

Figure 3 shows a flow chart of a comparison method for comparing images using the model images created by the present invention.

Figure 4 shows the results of the first comparative experiment of the present invention and known methods.

Figure 5 shows the results of a second comparative experiment of the present invention and a known method.

Figure 6 shows the results of a third comparative experiment of the present invention and known methods.

Claims (9)

A method for establishing a model for image recognition includes the following steps: inputting a plurality of digitized image files belonging to the n category, each image file including a pixel address representing the image and its grayscale data, and category information of the image; Most image files of each category Di(0<i n) extracting its common feature Ai; calculating a common feature Ai of a class and a common feature of a certain number (m) of other categories Aj (0<j m a difference characteristic n(Aj ^T Ai) of n-1, Dj≠Di); and adding the difference characteristic Σ(Aj ^T Ai) to the common feature Ai of the category to obtain a model image Ai* of the category; wherein the Ai * is calculated by a single step: stD _i =A _i +E _i . The method of claim 1, wherein m is n-1. For example, the method of claim 1 or 2 of the patent scope includes the steps of obtaining the model images A1* to An* of each category in the same manner. The method of claim 1 or 2, wherein the digitized image file in which the majority of the input points belong to the n category is a face digital image file belonging to n individuals. The method of claim 3, wherein the digitized image file in which the majority of the input points belong to the n category is a face digital image file belonging to n individuals. A method for determining a relationship between an input image and a majority of model images includes the following steps: inputting a digital image file to be processed, the image file including a pixel address representing an image and grayscale data thereof; Obtaining a majority of the model image obtained by the method of claim 1 or 2; comparing the difference between the image to be processed and the image of the majority of the model to calculate the similarity between the image to be processed and the model image of each category; The image to be processed is determined to belong to the category of the model image with the highest degree of similarity. A method for determining a relationship between an input image and a majority of model images includes the steps of: inputting a digital image file to be processed, the image file including a pixel address representing an image and grayscale data; obtaining a patent application scope The majority of the model image obtained by the third method; comparing the difference between the image to be processed and the image of the majority of the model to calculate the similarity between the image to be processed and the model image of each category; and determining that the image to be processed belongs to the The category of the most approximate model image. A method for determining a relationship between an input image and a majority of model images includes the steps of: inputting a digital image file to be processed, the image file including a pixel address representing an image and grayscale data; obtaining a patent application scope The majority model image obtained by the method of the first or second method; comparing the difference between the image to be processed and the image of the majority of the model to calculate the similarity between the image to be processed and the model image of each category; and determining the image to be processed It is a category of model images whose approximate approximation exceeds the critical value. A method for determining a relationship between an input image and a majority of model images includes the steps of: inputting a digital image file to be processed, the image file including a pixel address representing an image and grayscale data; obtaining a patent application scope Image of most categories of models obtained by the third method; Comparing the difference between the image to be processed and the image of the majority of the model to calculate the similarity between the image to be processed and the model image of each category; and determining that the image to be processed is a category of the model image whose approximate degree of approximation exceeds a critical value.