Image classification apparatus and method

Abstract

The present application provides an image classification device and method. The image classification device comprises a storage device, a calculation circuit and a classification circuit. The storage device can store information corresponding to a plurality of image categories. The calculation circuit is coupled to the storage device. The calculation circuit can obtain a target image from an image extraction device, and obtain a feature vector corresponding to the target image. In addition, the calculation circuit obtains a first estimation result corresponding to the target image according to the information corresponding to the plurality of image categories and the feature vector, and obtains a second estimation result corresponding to the target image according to a reference image, wherein the reference image corresponds to one of the plurality of image categories. The classification circuit is coupled to the calculation circuit. The classification circuit can add the target image to one of the plurality of image categories according to the first estimation result and the second estimation result.

Description

TECHNICAL FIELD

[0001] Embodiments of the present application generally relate to an image classification technique, and more particularly to an image classification technique that classifies an image according to a result generated by comparing a target image with a plurality of classified images and a result generated by a Hidden Markov Model (HMM) algorithm. BACKGROUND

[0002] With the advancement of technology, the application of image classification is becoming more and more popular. Therefore, how to generate more accurate image classification results is a topic worth studying. SUMMARY

[0003] In view of the problems of the prior art, embodiments of the present application provide an image classification device and method.

[0004] According to an embodiment of the present application, an image classification device is provided. The image classification device includes a storage device, a calculation circuit and a classification circuit. The storage device can be used to store information corresponding to a plurality of image categories. The calculation circuit is coupled to the storage device. The calculation circuit can obtain a target image from an image extraction device, and obtain a feature vector corresponding to the target image. In addition, the calculation circuit can obtain a first estimation result corresponding to the target image according to the information corresponding to the plurality of image categories and the feature vector, and obtain a second estimation result corresponding to the target image according to a reference image, wherein the reference image corresponds to one of the plurality of image categories. The classification circuit is coupled to the calculation circuit. The classification circuit can add the target image to one of the plurality of image categories according to the first estimation result and the second estimation result.

[0005] In some embodiments, each of the image categories includes a plurality of group images. In some embodiments, the calculation circuit can calculate a shortest distance of the feature vector corresponding to each of the image categories according to the feature vector and a group center of each of the group images of each of the image categories. When a minimum value of the shortest distances of the feature vector corresponding to each of the image categories is greater than a threshold value, the calculation circuit discards the target image. When the minimum value of the shortest distances of the feature vector corresponding to each of the image categories is not greater than a threshold value, the calculation circuit calculates the first estimation result according to the shortest distances of the feature vector corresponding to each of the image categories and a probability distribution algorithm.

[0006] In some embodiments, the classification circuit can multiply the first estimation result and the second estimation result to obtain a third estimation result, and add the target image to one of the plurality of image categories according to the third estimation result.

[0007] In some embodiments, the classification circuit can multiply the first estimation result by a first weight value to generate a first result, multiply the second estimation result by a second weight value to generate a second result, and add the first result and the second result to generate a third estimation result, and add the target image to one of the image categories according to the third estimation result.

[0008] In some embodiments, after the classification circuit adds the target image to one of the image categories, the classification circuit can update the information corresponding to the image category to which the target image is added.

[0009] According to an embodiment of the present application, an image classification method is provided. The image classification method is applicable to an image classification device. The image classification method comprises the following steps: obtaining a target image from an image extraction device; obtaining, by an operation circuit of the image classification device, a feature vector corresponding to the target image; obtaining, by the operation circuit, a first estimation result corresponding to the target image according to the information corresponding to the image categories and the feature vector; obtaining, by the operation circuit, a second estimation result corresponding to the target image according to a reference image, wherein the reference image corresponds to one of the image categories; and adding, by a classification circuit of the image classification device, the target image to one of the image categories according to the first estimation result and the second estimation result.

[0010] As to other additional features and advantages of the present application, those skilled in the art will find that the image classification device and method disclosed in the present application can be modified and improved without departing from the spirit and scope of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0011] Figure 1 FIG. 1 is a block diagram showing an image classification device 100 according to an embodiment of the present application.

[0012] Figure 2 FIG. 3 is a flow chart showing an image classification method according to an embodiment of the present application.

[0013] SYMBOL DESCRIPTION

[0014] 100: image classification device

[0015] 110: storage device

[0016] 120: operation circuit

[0017] 130: classification circuit

[0018] 200: image extraction unit

[0019] S210-S250: Steps DETAILED DESCRIPTION

[0020] The present application relates to an image classification device and method, and more particularly, to an image classification device and method for classifying images of different regions in a mouth.

[0021] Figure 1 is a block diagram showing an image classification device 100 according to an embodiment of the present application. As shown in Figure 1 , the image classification device 100 can include a storage device 110, a calculation circuit 120, and a classification circuit 130. Note that the block diagram shown in Figure 1 is merely for the convenience of explaining the embodiments of the present application, and the present application is not limited thereto. Other elements or other connection manners can also be included in the image classification device 100. According to an embodiment of the present application, the calculation circuit 120 and the classification circuit 130 can be integrated in a chip. Even more, the storage device 110, the calculation circuit 120, and the classification circuit 130 can also be integrated in a chip. Figure 1 According to an embodiment of the present application, the storage device 110 can be a volatile memory (e.g., a random access memory (RAM)), a non-volatile memory (e.g., a flash memory, a read only memory (ROM)), a hard disk, or a combination thereof. The storage unit can be used to store files and data required for image classification.

[0022] According to an embodiment of the present application, the storage device 110 can pre-store images of multiple image classes. According to an embodiment of the present application, each image class can correspond to a region, for example, different regions in a mouth, but the present application is not limited thereto. In addition, according to an embodiment of the present application, the images included in each image class can be pre-divided into multiple group images via a clustering algorithm.

[0023]

[0024] ​According to an embodiment of the present disclosure, the clustering algorithm can be a k-means algorithm, but the present disclosure is not limited thereto. In the k-means algorithm, the user can pre-determine the number of clusters into which the images included in an image category are to be divided (i.e., pre-determine the number of clusters included in each image category). In addition, in the k-means algorithm, each cluster corresponds to a cluster center in the k-means algorithm. That is, the number of cluster centers is the same as the number of clusters.

[0025] According to an embodiment of the present disclosure, the image classification device 100 can obtain a target image (i.e., an image extracted by the image extraction device 200 at the current time point) from the image extraction device 200 and add the target image to a proper image category. In an embodiment of the present disclosure, the image extraction device 200 can be moved to different positions at different time points to extract images corresponding to different image categories (regions). How to add the target image to the proper image category will be described in more detail below.

[0026] According to an embodiment of the present disclosure, when the target image is to be added to the proper image category, the operation circuit 120 can first extract a feature vector of the target image. In an embodiment, the operation circuit 120 can input the target image into a deep learning algorithm model to obtain a feature vector corresponding to the target image. In this embodiment, the deep learning algorithm can be a convolutional neural network (CNN) algorithm (e.g., MobileNet, EfficiebtNet, ResNet, etc.), but the present disclosure is not limited thereto.

[0027] After the operation circuit 120 obtains the feature vector corresponding to the target image, the operation circuit 120 calculates the distances between the feature vector corresponding to the target image and the cluster centers of each cluster of each image category, and obtains a shortest distance between the feature vector corresponding to the target image and each image category. For example, if there are N image categories stored in the storage device 110, and the images included in each image category are divided into M clusters, and the feature vector corresponding to the target image is v, then the shortest distance between the feature vector v corresponding to the target image and each image category can be represented as:

[0028] D v，n =min(d v，1 ,d v，2 ,…,d v，m ),n∈{1,2,…,N},m∈{1,2,…,M},

[0029] wherein D v，n represents the shortest distance between the feature vector v corresponding to the target image and the nth image category, d v，mrepresents the distance between the feature vector v corresponding to the target image and the cluster center of each cluster of image classes.

[0030] According to an embodiment of the present application, the operation circuit 120 determines whether the minimum value of the shortest distances between the feature vector corresponding to the target image and each image class is greater than a threshold value.

[0031] According to an embodiment of the present application, the threshold value can be calculated according to an interquartile range (IQR) algorithm, but the present application is not limited thereto. Specifically, the distances between all images stored in the storage device 110 and the cluster centers corresponding thereto are calculated in advance. Then, the distances between all images and the cluster centers corresponding thereto are counted according to the IQR algorithm to obtain the first quartile Q1, the second quartile Q2, and the third quartile Q3, and the interquartile range IQR (i.e., IQR = Q3 - Q1) in the IQR algorithm, and the threshold value is set to (Q3 + 1.5IQR).

[0032] If the minimum value of the shortest distances between the feature vector corresponding to the target image and each image class is greater than the threshold value (which means that the target image is not similar to each image class), the operation circuit 120 determines that the target image is a defective image and discards the target image. In the above example, if min(D v，1 , D v，2 , …, D v，N ) > threshold value, the operation circuit 120 determines to discard the target image.

[0033] If the minimum value of the shortest distances between the feature vector corresponding to the target image and each image class is not greater than the threshold value, the operation circuit 120 can perform a probability distribution operation on the inverses of the shortest distances between the feature vector corresponding to the target image and each image class according to a probability distribution algorithm to obtain the probabilities of each class corresponding to the target image (i.e., the first estimation result).

[0034] According to an embodiment of the present application, the probability distribution algorithm can be a softmax algorithm, but the present application is not limited thereto. The softmax algorithm can convert multi-class values into a probability distribution with a range of [0, 1] and a sum of 1. In the above example, the probabilities of each class corresponding to the target image (i.e., the first estimation result) can be represented as:

[0035]

[0036] where {p i，1 , p i，2 , …, p i，N} represents the probability of each class n (n ∈ {1, 2, …, N}) corresponding to the target image i.

[0037] According to an embodiment of the present application, the operation circuit 120 also calculates the probability of each category corresponding to the target image (i.e., the second estimation result) according to a Hidden Markov Model (HMM) algorithm and a reference image. According to an embodiment of the present application, the reference image is the target image at a previous time point, and the image category corresponding to the reference image is one of the image categories stored in the storage device 110. For example, the reference image is the target image at time point t-1, and the image category (or region) corresponding to the reference image is known (estimated, as described above). Therefore, the reference image and the image category corresponding to the reference image can be used to estimate the probability of each image category corresponding to the target image at the current time point t (i.e., the probability of the image category (region) corresponding to the reference image moving to each image category (region) from time point t-1 (the reference image) to the current time point t (the target image) by the image extraction device 200).

[0038] According to an embodiment of the present application, the storage device 110 further stores the moving probability of each image category (region) corresponding to each other. Specifically, the moving probability of each region corresponding to each other represents the probability of the image extraction device 200 moving from one region to another region when the image extraction device 200 extracts images at consecutive time points. The operation circuit 120 can obtain the probability of each image category corresponding to the target image (i.e., the second estimation result) according to the moving probability of each region corresponding to each other and the image category (region) corresponding to the reference image by the HMM algorithm.

[0039] According to another embodiment of the present application, the storage device 110 further stores distance information of each image category (region) corresponding to each other. More specifically, the distance Z i，j of each image category (region) corresponding to each other is measured in advance and recorded in the storage device 110, where Z i，j represents the distance of moving from region i to region j. In addition, in this embodiment, the operation circuit 120 can obtain a displacement information between the target image and the reference image according to a first algorithm.

[0040] In an embodiment, the first algorithm can be an image comparison algorithm (e.g., a feature point comparison algorithm). In this embodiment, the operation circuit 120 can compare the feature points included in the target image and the reference image, and calculate how many pixels the feature points have moved. After obtaining the moved pixels, the operation circuit 120 can estimate the distance change d (i.e., the displacement information) between the target image and the reference image since the pixels and the distance have a proportional relationship.

[0041] In another embodiment, the first algorithm can be an inertial measurement unit (IMU) algorithm. In this embodiment, the operation circuit 120 can obtain the acceleration change and the time change of the reference image and the target image according to the inertial measurement unit (IMU) information of the reference image and the inertial measurement unit (IMU) information of the target image, and then estimate the distance change d (i.e., the displacement information) between the reference image and the target image.

[0042] After obtaining the displacement information, the operation circuit 120 can obtain the probability of each image class (region) of the target image (i.e., the second estimation result) according to the distance information (i.e., the distance Z i，j ) corresponding to each image class (region) of the target image, the image class (region) corresponding to the reference image (for example, region i), and the calculated displacement information (i.e., the distance change d) by using the HMM algorithm. Specifically, the operation circuit 120 can bring the difference (i.e., |d-Z i，j ) between the distance change d and the distance Z i，j into a probability density function to generate a probability density function distribution diagram (i.e., the second estimation result). For example, the operation circuit 120 can bring the difference (i.e., |d-Z i，j ) between the distance change d and the distance Z i，j into an exponential distribution function:

[0043]

[0044] wherein the operation circuit 120 takes the difference (i.e., |d-Z i，j ) between the distance change d and the distance Z i，j as the input value x of the exponential distribution function. In this embodiment, the smaller the difference (i.e., |d-Z i，j ) between the distance change d and the distance Z i，j , the greater the probability value of moving from the region i (i.e., the region corresponding to the reference image) to the corresponding region j (i.e., the region to which the target image can correspond). The operation circuit 120 can estimate the probability of each image class (region) of the target image (i.e., the second estimation result) by using this embodiment.

[0045] According to another embodiment of the present application, the storage device 110 further stores angle information corresponding to each image class (region) to each other. More specifically, the angle r i，j corresponding to each image class (region) is measured in advance and recorded in the storage device 110, wherein r i，j represents the angle of moving from the region i to the region j. In addition, in this embodiment, the operation circuit 120 obtains an angle change information between the target image and the reference image according to a second algorithm.

[0046] In one embodiment, the second algorithm can be an inertial measurement unit algorithm. In this embodiment, the operation circuit 120 can obtain the rotation angle change c (i.e., the angle change information) between the reference image and the target image according to the inertial measurement unit (IMU) information of the reference image and the inertial measurement unit (IMU) information of the target image.

[0047] After obtaining the angle change information, the operation circuit 120 can obtain the probability of each image class (region) of the target image (i.e., the second estimation result) according to the angle information (i.e., the angle r i，j ) corresponding to each other of each image class (region), the image class (region) corresponding to the reference image (for example, the region i), and the angle change information (i.e., the rotation angle change c) by the HMM algorithm. Specifically, the operation circuit 120 can bring the difference (i.e., |c-r i，j ) between the rotation angle change c and the angle r i，j into a probability density function to generate a probability density function distribution (i.e., the second position estimation result). For example, the operation circuit 120 can bring the difference (i.e., |c-r i，j ) between the rotation angle change c and the angle r i，j into an exponential distribution function:

[0048]

[0049] wherein the operation circuit 120 takes the difference (i.e., |c-r i，j ) between the rotation angle change c and the angle r i，j as the input value x of the exponential distribution function. In this embodiment, the smaller the difference (i.e., |c-r i，j ) between the rotation angle change c and the angle r i，j , the greater the probability value of moving from the region i (i.e., the region corresponding to the reference image) to the corresponding region j (i.e., the region to which the target image can correspond). The operation circuit 120 can estimate the probability of each image class of the target image (i.e., the second estimation result) by this embodiment.

[0050] According to one embodiment of the present application, when the operation circuit 120 obtains the first estimation result and the second estimation result, the classification circuit 130 can multiply the first estimation result and the second estimation result to obtain a third estimation result. Then, the classification circuit 130 can add the target image into the image class corresponding to the maximum value in the third estimation result (i.e., the image class to which the target image is most likely to correspond) according to the maximum value in the third estimation result. For example, if the first estimation result {p i，1 , p i，2 , …, p i，N} and the second estimation result {h i，1 , hi，2 ,..., h i，N}, the third estimation results and the maximum value among the third estimation results can be respectively represented as:

[0051] P i，n = p i，n × h i，n , n e {1, 2,..., N},

[0052] C i = argmax(P i，1 , P i，2 ,..., P i，N )

[0053] where P i，n represents the third estimation result corresponding to the target image i, and C i represents the maximum value among the third estimation results.

[0054] According to another embodiment of the present application, when the operation circuit 120 obtains the first estimation results and the second estimation results, the classification circuit 130 can multiply the first estimation results by a first weight value to generate a first result, and multiply the second estimation results by a second weight value to generate a second result. Then, the classification circuit 130 can add the first result and the second result to generate a third estimation result. Finally, the classification circuit 130 can add the target image into the image category corresponding to the maximum value among the third estimation results (i.e. the image category to which the target image is most likely to correspond) according to the maximum value among the third estimation results. For example, if the first estimation results are {p i，1 , p i，2 ,..., p i，N} and the second estimation results are {h i，1 , h i，2 ,..., h i，N}, the third estimation results and the maximum value among the third estimation results can be respectively represented as:

[0055] P i，n = w1 x p i，n + w2 x h i，n , n e {1, 2,..., N},

[0056] C i = argmax(P i，1 , P i，2 ,..., P i，N )

[0057] where P i，n represents the third estimation result corresponding to the target image i, w1 represents the first weight value, w2 represents the second weight value, and C i represents the maximum value among the third estimation results.

[0058] According to an embodiment of the present application, after the classification circuit 130 adds the target image into the image category corresponding to the maximum value of the third estimation result, the classification circuit 130 can further update the information corresponding to the image category to which the target image is added by using a clustering algorithm (for example, the k-means algorithm). For example, after the target image is added into an image category, the cluster center of each cluster included in the image category can be changed so as to more accurately determine the image category corresponding to the target image at the next time point or the next image.

[0059] Figure 2 is a flowchart of an image classification method according to an embodiment of the present application. The image classification method can be applied to the image classification device 100. As shown in Figure 2 At step S210, the image classification device 100 obtains a target image from an image extraction unit.

[0060] At step S220, the operation circuit of the image classification device 100 obtains a feature vector corresponding to the target image.

[0061] At step S230, the operation circuit of the image classification device 100 obtains a first estimation result corresponding to the target image according to the information corresponding to the plurality of image categories and the feature vector corresponding to the target image.

[0062] At step S240, the operation circuit of the image classification device 100 obtains a second estimation result corresponding to the target image according to a Hidden Markov Model (HMM) algorithm and a reference image, wherein the reference image corresponds to one of the plurality of image categories.

[0063] At step S250, the classification circuit of the image classification device 100 adds the target image into one of the plurality of image categories according to the first estimation result and the second estimation result.

[0064] According to an embodiment of the present application, in the image classification method, each image category can include a plurality of cluster images.

[0065] According to an embodiment of the present application, at step 430 of the image classification method, the operation circuit of the image classification device 100 can calculate the shortest distance of the feature vector corresponding to the target image to each image category according to the feature vector corresponding to the target image and the cluster center of each cluster image of each image category.

[0066] According to an embodiment of the present application, in the image classification method, when the minimum value of the shortest distances of the feature vector corresponding to the target image to each image category is greater than a threshold value, the operation circuit of the image classification device 100 can discard the target image. When the minimum value of the shortest distances of the feature vector corresponding to the target image to each image category is not greater than the threshold value, the operation circuit of the image classification device 100 calculates a first estimation result according to the shortest distances of the feature vector corresponding to the target image to each image category and a probability distribution algorithm.

[0067] According to an embodiment of the present application, in step 450 of the image classification method, the classification circuit of the image classification device 100 can multiply the first estimation result and the second estimation result to obtain a third estimation result, and add the target image to one of the image categories according to the third estimation result.

[0068] According to another embodiment of the present application, in step 450 of the image classification method, the classification circuit of the image classification device 100 can multiply the first estimation result by a first weight value to generate a first result, and multiply the second estimation result by a second weight value to generate a second result. Then, the classification circuit of the image classification device 100 can add the first result and the second result to generate a third estimation result, and add the target image to one of the image categories according to the third estimation result.

[0069] According to an embodiment of the present application, in the image classification method, after adding the target image to one of the image categories, the classification circuit of the image classification device 100 can update the information corresponding to the image category to which the target image is added.

[0070] The image classification method according to the present application can more accurately obtain the image category corresponding to the target image by the image classification device. In addition, the image classification method according to the present application can continuously update the information corresponding to each image category by the image classification device, so as to increase the accuracy of subsequent image classification.

[0071] The sequence numbers in the specification and claims, such as "first", "second", etc., are only for the convenience of description and do not have any order relationship between each other.

[0072] The steps of a method or algorithm described in connection with the present disclosure can be embodied directly in hardware, in a software module executed by a processor, or in a combination of the two. A software module (including an instruction set) can reside in data storage media, RAM, flash memory, ROM, EPROM, EEPROM, registers, hard disk, portable disk, CD-ROM, DVD, or any other form of computer-readable storage medium known in the art. A memory medium can be coupled to a machine device, such as a computer / processor (for convenience, referred to herein as a processor), which can read information (such as program code) from and write information to the memory medium. A memory medium can be integrated into a processor. A special- purpose application-specific integrated circuit (ASIC) comprising a processor and a memory medium can be included in a user device. In other words, a processor and a memory medium are included in a user device without being directly connected to the user device. In addition, in some embodiments, any suitable computer program product includes a readable storage medium readable and / or writable by a machine, including one or more of the disclosed embodiments, program code embodied in the medium for use with a processor. In some embodiments, program code embodied on a computer program product can include packaging material.

[0073] The above-described paragraphs describe multiple aspects. It is clear that the teachings herein can be implemented in various ways and that any particular architectural or functional details are only examples of representative systems rather than limiting the present disclosure. According to the present teachings, a person of ordinary skill in the art will understand that each of the aspects disclosed herein can be implemented independently or in combination with one or more other aspects.

[0074] While the present disclosure has been disclosed in connection with the embodiments presented, it should be understood that certain modifications can be made to the described embodiments and additional implementations can be implemented without departing from the spirit and scope of the disclosure. Therefore, other implementations are within the scope of the following claims.

Claims

1. An image classification apparatus comprising: a storage device storing information corresponding to a plurality of image classes; and an operation circuit coupled to the storage device, the operation circuit obtaining a target image from an image extraction device and obtaining a feature vector corresponding to the target image, wherein the operation circuit obtains a first estimation result corresponding to the target image based on the information corresponding to the plurality of image classes and the feature vector, and wherein the operation circuit obtains a second estimation result corresponding to the target image based on a reference image, wherein the reference image corresponds to one of the plurality of image classes; and a classification circuit coupled to the operation circuit, wherein the classification circuit adds the target image to one of the plurality of image classes based on the first estimation result and the second estimation result, wherein each of the image classes comprises a plurality of cluster images, and wherein the operation circuit calculates a distance between the feature vector and a cluster center of each of the cluster images of each of the image classes based on the feature vector and the cluster center of each of the cluster images of each of the image classes, and obtains a shortest distance of the feature vector corresponding to each of the image classes, the shortest distance being represented as: D v,n = min(d v,1 ,d v,2 ,…,d v,m ), n e {1,2,…,N}, m e {1,2,…,M}, where N represents the number of image categories, M represents the number of groups contained in each image category, d v,m represents the distance between the feature vector v corresponding to the target image and the group center of each group, D v,n represents the shortest distance between the feature vector v corresponding to the target image and the n-th image category. wherein when a minimum value of the shortest distance of the feature vector corresponding to each of the image classes is not greater than a threshold value, the operation circuit calculates the first estimation result based on the shortest distance of the feature vector corresponding to each of the image classes and a probability distribution algorithm, and wherein when the minimum value of the shortest distance of the feature vector corresponding to each of the image classes is greater than the threshold value, the operation circuit discards the target image.

2. The image classification apparatus of claim 1, wherein the classification circuit multiplies the first estimation result and the second estimation result to obtain a third estimation result, and adds the target image to one of the plurality of image classes based on the third estimation result.

3. The image classification apparatus of claim 1, wherein the classification circuit multiplies the first estimation result by a first weight value to generate a first result, and multiplies the second estimation result by a second weight value to generate a second result, and the classification circuit adds the first result and the second result to generate a third estimation result, and adds the target image to one of the plurality of image classes based on the third estimation result.

4. The image classification apparatus of claim 1, wherein the classification circuit updates the information corresponding to the image class to which the target image is added after the target image is added to one of the plurality of image classes.

5. An image classification method suitable for an image classification apparatus, comprising: obtaining a target image from an image extraction device; obtaining a feature vector corresponding to the target image by an operation circuit of the image classification apparatus; obtaining a first estimation result corresponding to the target image based on information corresponding to a plurality of image classes and the feature vector by the operation circuit; obtaining a second estimation result corresponding to the target image based on a reference image by the operation circuit, wherein the reference image corresponds to one of the plurality of image classes; and adding the target image into one of the plurality of image categories according to the first estimation result and the second estimation result by a classification circuit of the image classification device, wherein each of the image categories comprises a plurality of group images, and The image classification method further comprises: calculating distances between the feature vector and a group center of each group of each of the image categories according to the feature vector and the group center of each group of each of the image categories by the operation circuit, and obtaining a shortest distance of the feature vector corresponding to each of the image categories, the shortest distance being represented as: D v,n = min(d v,1 ,d v,2 ,…,d v,m ), n e {1,2,…,N}, m e {1,2,…,M}, wherein N represents the number of image categories, M represents the number of groups contained in each image category, d v,m represents the distance between the feature vector v corresponding to the target image and the group center of each group, D v,n represents the shortest distance between the feature vector v corresponding to the target image and the n-th image category. calculating the first estimation result according to the shortest distance of the feature vector corresponding to each of the image categories and a probability distribution algorithm by the operation circuit when a minimum value of the shortest distance of the feature vector corresponding to each of the image categories is not greater than a threshold value; and discarding the target image by the operation circuit when the minimum value of the shortest distance of the feature vector corresponding to each of the image categories is greater than the threshold value.

6. The image classification method of claim 5, further comprising: multiplying the first estimation result and the second estimation result by the classification circuit to obtain a third estimation result, and adding the target image into one of the plurality of image categories according to the third estimation result.

7. The image classification method of claim 5, further comprising: multiplying the first estimation result by a first weight value to generate a first result and multiplying the second estimation result by a second weight value to generate a second result by the classification circuit, and adding the first result and the second result by the classification circuit to generate a third estimation result, and adding the target image into one of the plurality of image categories according to the third estimation result.

8. The image classification method of claim 5, further comprising: updating the information corresponding to the image category to which the target image is added after the target image is added into one of the plurality of image categories by the classification circuit.

Images