Image processing method, device, computer equipment and storage medium

Abstract

The present invention relates to the field of artificial intelligence technology, and discloses an image processing method, device, computer equipment and storage medium, the method comprising: obtaining multiple images to be processed; performing portrait clustering processing on the multiple images to be processed to obtain multiple portrait cluster sets containing the same portrait; performing static portrait analysis on each portrait cluster set to obtain analysis results corresponding to each portrait cluster set; based on the analysis results corresponding to each portrait cluster set, performing mis-collection filtering on each portrait cluster set to generate image acquisition results for each portrait cluster set. Therefore, the present invention avoids the interference of static portraits and improves the accuracy and correctness of image processing of real portraits.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an image processing method, device, computer equipment and storage medium. Background Art

[0002] With the development of computer technology and artificial intelligence, the means of scientific and technological informationization are constantly being innovated, especially the development of facial technology, which is gradually applied to various fields in people's daily life, providing strong support for the safety or convenience of people's daily life. Therefore, a big data platform for portrait data collection and recognition is built. However, in the portrait collection area of ​​portrait data collection, there will be many static portraits such as print advertisements or posters in daily life, which will become interference items for image processing or subsequent portrait recognition, resulting in continuous miscollection of such static portraits, which greatly affects the accuracy and correctness of image processing. Moreover, since such static portraits have no effect in the image processing process, the pressure on data storage and processing of image processing is increased. Summary of the invention

[0003] The present invention provides an image processing method, an apparatus, a computer device and a storage medium, which can automatically analyze and filter out mistakenly captured portrait images from a captured video to be recognized, and only perform face recognition on portrait images containing real portraits, thereby avoiding interference from static portraits and improving the accuracy and correctness of image processing in the captured video.

[0004] An image processing method, comprising:

[0005] Multiple images to be processed;

[0006] Performing portrait clustering processing on the plurality of images to be processed to obtain a plurality of portrait cluster sets containing the same portrait;

[0007] Performing static portrait analysis on each of the portrait cluster sets to obtain analysis results corresponding to each of the portrait cluster sets;

[0008] Based on the analysis results corresponding to each of the portrait cluster sets, mis-capture filtering is performed on each of the portrait cluster sets to generate image capture results for each of the portrait cluster sets.

[0009] An image processing device, comprising:

[0010] An acquisition module, used for acquiring a plurality of images to be processed;

[0011] A clustering module, used for performing portrait clustering processing on the plurality of images to be processed to obtain a plurality of portrait cluster sets containing the same portrait;

[0012] An analysis module, used to perform static portrait analysis on each of the portrait cluster sets to obtain analysis results corresponding to each of the portrait cluster sets;

[0013] The filtering module is used to filter out mis-collected images for each of the portrait cluster sets based on the analysis results corresponding to each of the portrait cluster sets, and generate image collection results for each of the portrait cluster sets.

[0014] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above-mentioned image processing method when executing the computer program.

[0015] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the above-mentioned image processing method are implemented.

[0016] The image processing method, device, computer equipment and storage medium provided by the present invention obtain multiple images to be processed; perform portrait clustering processing on the multiple images to be processed to obtain multiple portrait cluster sets containing the same portrait; perform static portrait analysis on each of the portrait cluster sets to obtain analysis results corresponding to each of the portrait cluster sets; based on the analysis results corresponding to each of the portrait cluster sets, perform mis-collection filtering on each of the portrait cluster sets to generate image acquisition results for each of the portrait cluster sets. Therefore, the present invention automatically obtains images to be processed containing portraits, and automatically analyzes mis-collected portrait cluster sets through portrait clustering processing and static portrait analysis, and generates image acquisition results for each of the portrait cluster sets after mis-collection filtering, thereby avoiding interference from static portraits, reducing subsequent image processing operations on static portrait images, and improving the accuracy and correctness of image processing of real portraits. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.

[0018] Figure 1 is a schematic diagram of an application environment of an image processing method according to an embodiment of the present invention;

[0019] Figure 2 is a flow chart of an image processing method in one embodiment of the present invention;

[0020] Figure 3 is a flow chart of step S10 of the image processing method in one embodiment of the present invention;

[0021] Figure 4 is a flow chart of step S20 of the image processing method in one embodiment of the present invention;

[0022] Figure 5 is a principle block diagram of an image processing device in one embodiment of the present invention;

[0023] Figure 6 is a schematic diagram of a computer device in one embodiment of the present invention. DETAILED DESCRIPTION

[0024] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0025] The image processing method provided by the present invention can be applied to Figure 1 In an application environment, a client (computer device or terminal) communicates with a server via a network. The client (computer device or terminal) includes but is not limited to various personal computers, laptops, smart phones, tablet computers, and portable wearable devices. The server can be implemented as an independent server or a server cluster consisting of multiple servers.

[0026] In one embodiment, if Figure 2 As shown, an image processing method is provided, and its technical solution mainly includes the following steps S10-S40:

[0027] S10, obtaining a plurality of images to be processed.

[0028] It can be understood that the image to be processed is an image containing a portrait that needs to be identified as to whether it is an erroneous image. The process of acquiring the image to be processed can be set according to the needs. The process of acquiring the image to be processed can be to perform single portrait detection on each candidate image from multiple candidate images, obtain the detection results of each candidate image, and determine the candidate image corresponding to the detection result containing the portrait as the acquisition process of the image to be processed.

[0029] The candidate image can be an image taken by an image acquisition device or an image stored in a preset database. The process of single portrait detection is to extract the portrait features of a single portrait from the input image, identify the image based on the extracted portrait features, and identify the image containing a portrait.

[0030] Portrait features are features related to the outline of the human body, such as the face, eyes, nose, mouth, ears, eyebrows, shoulders, arms and other features related to the upper body or the outline of a portrait. Identifying an image containing a portrait includes, when multiple portraits are identified, comparing the size and area of ​​the identified portrait areas or the proportion of the input image, obtaining the identified portrait corresponding to the portrait area with the largest size and area or the largest proportion as the target portrait of the input image, and in the subsequent portrait clustering and mis-acceptance analysis process, only the target portrait is processed.

[0031] Optionally, the process of obtaining the image to be processed can also be a video to be identified containing a portrait shot by a video acquisition device of a big data platform. The duration of the video to be identified can be set according to demand. For example, the duration can be set to 1 second = 12 frames, 3 seconds = 36 frames, etc. The method of obtaining the video to be identified can be to detect the presence of a portrait in real time, thereby obtaining the preset duration in real time, or to detect the portrait from the stored video content and extract the video with the portrait, divide the video into preset durations, and thereby obtain the divided video.

[0032] The process of detecting a portrait is to extract the features of human body parts from each frame of the captured video through the trained target detection model, identify the frame based on the extracted features of human body parts, and identify the recognition result of the frame. When the recognition result is detected to represent a portrait with human body parts, it is determined that the frame has a portrait, indicating the process of detecting the portrait, and the extraction can be started from this frame until a frame without a portrait is detected, and then the video with the portrait is extracted. The video to be recognized is framed, and the frame processing is the operation process of segmenting or extracting a frame of image from each frame of the input video, thereby obtaining multiple images to be processed. The features of human body parts are features related to the parts of the human body, such as the features of human body parts are head, hands, feet, shoulders, etc.

[0033] The training process of the target detection model inputs a sample image containing a human body part label to be trained, extracts human body part features from the sample image through the model containing initial parameters, and classifies the sample image according to the extracted human body part features, classifies the sample image containing human body parts, calculates the loss value between the sample recognition result and the human body part label associated with the sample image, and iteratively updates the initial parameters of the target detection model according to the loss value until the loss value reaches a convergence condition, wherein the convergence condition may be that the loss value no longer changes or the number of iterations reaches a preset number, stops training, and records the target detection model after convergence as the trained target detection model.

[0034] In one embodiment, if Figure 3As shown, in step S10, a plurality of images to be processed are obtained, including:

[0035] S101, obtaining multiple candidate images.

[0036] Understandably, the candidate images may be images captured by an image acquisition device, or images stored in a preset database, that is, a large amount of captured images are stored in a database of a preset path or a cloud database, and the candidate images are obtained from the database of the preset path or the cloud database. The candidate images may also be videos captured in real time by a video acquisition device, and the captured videos are subjected to frame processing, which is the process of segmenting each frame of the input video or extracting a frame of image, thereby obtaining multiple candidate images.

[0037] S102, performing single person detection on each candidate image to obtain a detection result of each candidate image.

[0038] Understandably, single portrait detection is a detection process that can detect a single portrait. The single portrait detection process for each candidate image can be implemented through a trained target recognition model. The candidate image is input into the target recognition model, and the candidate image is transformed and convolved through the target recognition model to extract portrait features. Portrait features are features related to the outline of the human body, such as face, eyes, nose, mouth, ears, eyebrows, shoulders, arms, etc. related to the upper body of the human body or the outline of a portrait. The network structure of the target recognition model can be set according to demand. For example, the network structure of the target recognition model can be CenterNet, Faster R-CNN, SSD, YOLO and other network structures.

[0039] Preferably, the network structure of the target recognition model is the network structure of CenterNet, and the processing process of the target recognition model is to preprocess the input candidate image. First, the candidate image needs to be scaled to a preset size (e.g., 512×512). The preprocessing includes image scaling, that is, using image scaling technology, scaling the long and short sides of the image to a preset size by scaling the long side and padding the short side with zeros; secondly, performing image channel separation on the scaled candidate image to separate the image to be processed containing three channels. Image channel separation is to separate the red, green, and blue channels in the image, thereby completing the preprocessing of the input candidate image and obtaining preprocessed data.

[0040] The portrait features of the preprocessed data are extracted through the ResNet50 network in the target recognition model based on CenterNet. The target recognition model detects the portrait as a target point, that is, the center point of the target area is used to represent the portrait, and the center point offset (offset), width and height (size) of the portrait are predicted to obtain the actual area of ​​the object; then the extracted portrait features are upsampled by the deconvolution module, that is, deconvolution (the inverse process of convolution), and upsampling can be performed three times to obtain the feature map to be processed. Finally, the feature map to be processed is predicted by three branch networks, namely the heat map prediction network, the length and width prediction network and the regional center offset prediction network. The detection network is used to classify the feature map to be processed through the heat map prediction network, and the center point of the target area of ​​each portrait and the radius of the Gaussian circle are predicted. With the center point as the center of the circle, the radius is decreased outward according to the Gaussian function to obtain the heat map corresponding to the feature map to be processed; the length and width area of ​​multiple targets are predicted through the length and width prediction network to obtain the length and width map of the portrait corresponding to the feature map to be processed, and the offset value of the target is predicted through the regional center offset prediction network to obtain the center offset value of the target area of ​​the portrait corresponding to the feature map to be processed. According to the heat map, the length and width map and the center offset value, the area of ​​the portrait is determined, that is, the detection result of the portrait in the candidate image.

[0041] The detection result reflects the portrait area (region range) with portrait features in the candidate image, and determines that the detection result contains a portrait. In this way, the target recognition model can transform the high-dimensional features into low-dimensional feature representations through the feature space, and then perform classification and recognition through the low-dimensional feature representation to identify the detection result of whether the region range contains a portrait.

[0042] In one embodiment, the network structure of the target recognition model is the network structure of YOLO V3, because the network structure of YOLO V3 can quickly locate the position of the human body from different scales. The training process of the target recognition model is to input a human sample image containing a portrait annotated area, extract the portrait features (convolution) of the human sample image, obtain a small-scale feature map based on the extracted portrait features, upsample and tensor splice the small-scale feature map to obtain a medium-scale feature map, and then upsample and tensor splice the medium-scale feature map to obtain a large-scale feature map.

[0043] Furthermore, multi-scale detection is performed on the small-scale, medium-scale and large-scale feature maps to predict a rectangular area of ​​a portrait with portrait features, calculate the loss value between the area and the portrait annotation area, and iteratively update the parameters of the target recognition model according to the loss value until the loss value reaches the convergence condition, thereby stopping the training process. The small scale, medium scale and large scale can be 1 / 32 scale, 1 / 16 scale and 1 / 8 scale relative to the candidate image size, respectively. Feature maps of different scales can detect more fine-grained portrait features. Upsampling is the reverse propagation of convolution, that is, the transpose of deconvolution or forward propagation. Tensor splicing is the process of splicing the upsampling results of the intermediate layer in the convolution process and the adjacent next layer. Portrait features are features related to human body parts, such as head, hair, hands, face, torso, clothes, etc. that can reflect the outline of a portrait. The process of multi-target single portrait detection is to extract portrait features of the input candidate image through the target recognition model, predict the human body area based on the extracted portrait features, predict a rectangular area containing a portrait, and determine that the detection result contains a portrait.

[0044] S103: Determine the candidate image corresponding to the detection result containing the portrait as the image to be processed.

[0045] It can be understood that the candidate image corresponding to the detection result containing the portrait is recorded as the image to be processed. When multiple portraits are identified, the size and area of ​​the identified portrait areas are compared or compared with the proportion of the input image, and the identified portrait corresponding to the portrait area with the largest size and area or the largest proportion is obtained as the target portrait of the input image. That is, the detection result reflects the portrait area with the largest size and area or the largest proportion in the candidate image. In the subsequent portrait clustering and mis-acceptance analysis process, only the portrait area of ​​the target portrait is processed.

[0046] The present invention achieves the following steps: acquiring a plurality of candidate images; performing single portrait detection on each candidate image to obtain a detection result of each candidate image; and determining the candidate image corresponding to the detection result containing the portrait as the image to be processed. In this way, the single portrait detection can be used to detect the image to be processed containing the portrait in the input candidate images, thereby reducing the process of manual portrait recognition and improving the accuracy and efficiency of image processing.

[0047] S20, performing portrait clustering processing on the multiple images to be processed to obtain multiple portrait cluster sets containing the same portrait.

[0048] Understandably, the process of performing portrait clustering processing on multiple images to be processed is as follows: first, extracting facial attribute features from each image to be processed, and determining a facial feature map of each image to be processed based on the extracted facial attribute features, wherein the facial attribute feature map reflects the information of the face of the portrait in the image to be processed or also contains features of other attributes, wherein other attributes include clothes, hair, neck, etc., and the facial attribute features are the face of the portrait or also contain features of other attributes; secondly, clustering processing is performed on all facial feature maps to obtain multiple portrait clusters, wherein the clustering processing process can be set according to requirements, for example, the clustering processing process can be performed using K- means algorithm searches for K portrait clusters for all facial feature maps. The clustering process can also be to compare the similarities between two portrait feature maps and aggregate the two portrait feature maps whose similarities reach a preset similarity threshold into one portrait cluster. Finally, according to the portrait clusters corresponding to each image to be processed, a process of obtaining multiple portrait cluster sets containing the same portrait is performed, that is, the images to be processed containing the same portrait are aggregated into one portrait cluster set, so that multiple portrait cluster sets can be obtained.

[0049] The facial attribute feature extraction process is to convert the image information in the high-dimensional image to be processed into low-dimensional image information through pixel convolution, and convolve the correlation between pixels during the conversion process, so as to extract the pixel information or feature information in the portrait image, thereby extracting the facial attribute features, and finally determine the facial feature map of each image to be processed.

[0050] In one embodiment, if Figure 4 As shown, in step S20, portrait clustering processing is performed on multiple images to be processed to obtain multiple portrait cluster sets containing the same portrait, including:

[0051] S201, extracting facial attribute features from each image to be processed to obtain a facial feature map of each image to be processed.

[0052] It can be understood that for an image to be processed, when extracting the facial attribute features of the image to be processed, structured transformation, grayscale processing, convolution processing, pooling and dimensionality reduction processing can be included.

[0053] First, the image to be processed needs to be structurally converted and grayscale processed to obtain channel information. Structural conversion is the process of converting an irregular or non-uniform unstructured picture / image into a digital format that can be transmitted or stored and processed by a computer, thereby converting it into digitally embodied structured image information.

[0054] The process of structural transformation of each image to be processed can be achieved through a feature extraction model. The feature extraction model is a trained neural network model for extracting feature information between pixels in an input image. The feature extraction model includes an input layer, a convolutional layer, and a fully connected layer. The process of structural transformation of the image to be processed is: inputting the image to be processed into the feature extraction model, performing image scaling processing using image scaling technology on the image to be processed through the input layer in the feature extraction model, scaling it to a preset size, and then performing digital conversion of the three-channel information of R (red), G (green), and B (blue) for the scaled portrait image, that is, performing digital conversion of the three-channel information of R (red), G (green), and B (blue) for each pixel in the scaled portrait image, and converting the scaled portrait image into a preset size. The portrait image is divided into red channel, green channel and blue channel, and then the pixel value of each pixel in red depth or strength is encoded by a value of 0 to 255 in the red channel, the pixel value of each pixel in green depth or strength is encoded by a value of 0 to 255 in the green channel, and the pixel value of each pixel in blue depth or strength is encoded by a value of 0 to 255 in the blue channel. Grayscale processing is a process of graying the pixel value matrix of the three channels (i.e., the image to be processed after the structured conversion), so as to obtain the channel information of a channel, also called a grayscale image, that is, the pixel value of the red channel, the pixel value of the green channel and the pixel value of the blue channel corresponding to the same pixel point are weighted and summed to obtain the pixel value of the gray channel corresponding to the pixel point, wherein the weight of each channel can be set according to the demand. Preferably, the weight of each channel is 0.3 for the red channel, 0.59 for the green channel and 0.11 for the blue channel.

[0055] Secondly, when performing convolution processing, the grayscale image is convolved with multiple levels of different convolution kernels through the convolution layer to extract facial attribute features. The facial attribute features are the features of the face of a portrait or also include other attributes. Among them, other attributes include clothes, hair, neck, etc., and pixel information or feature information can be extracted. The convolution layer outputs multiple channels of initial feature maps containing facial attribute features. The initial feature map reflects the pixel information or feature information of each pixel and between each pixel.

[0056] Finally, when performing pooling and dimensionality reduction processing, the multi-channel initial feature map output by the convolution layer is pooled and reduced in dimension through the fully connected layer, that is, the high-dimensional multi-channel feature map is reduced in dimension to a one-dimensional array. The pooling process is a global average pooling process, and the dimensionality reduction process is a convolution process to one dimension, so that a one-dimensional feature vector array can be obtained. The one-dimensional feature vector array is grouped and summarized to obtain a facial feature map. The facial feature map can be a numerical value, a two-dimensional coordinate value, or a three-dimensional value. The facial feature maps converted from the portrait images of the same portrait are equal or similar, because the portrait information contained in the portrait images is the same or similar.

[0057] S202: performing clustering processing on all face feature images to obtain multiple portrait clusters.

[0058] Understandably, the clustering process can be set according to demand. For example, the clustering process can be to use the K-means algorithm to find K portrait clusters for all facial feature maps. All facial feature maps are used as a data set, and the data set is input into a clustering model. The clustering model is a model for portrait clustering processing based on K-means. The K-means clustering algorithm is used through the clustering model to perform K-means-based portrait clustering processing on the input data set. The K-means-based portrait clustering processing process is the process of using the K-means clustering algorithm to classify K portrait clusters. Different portrait clusters represent different portraits. The K-means clustering algorithm randomly selects K points in a data set. These K points represent the center points of K clusters. Then, the distance between each predicted height value in the one-dimensional array and these K center points is calculated. A predicted height value is divided into the cluster corresponding to the center point closest to it. Finally, for the classification cluster just generated, the center point of the cluster is recalculated. The value of K can be increased or decreased during the calculation process until the algorithm converges to obtain a suitable result of K center points. The facial feature maps covered by a center point and close to it are combined into a cluster, and the cluster is determined as a portrait cluster. One portrait represents a portrait cluster.

[0059] Optionally, clustering processing can also be to compare the similarities between two portrait feature maps, such as using a cosine similarity algorithm to calculate the cosine value between two portrait feature maps, measure the degree of similarity between the two using the cosine values, obtain the similarity, and group two portrait feature maps whose similarity reaches a preset similarity threshold into a portrait cluster, thereby being able to group multiple similar portrait feature maps into a portrait cluster. In this way, in the case of multiple portraits, multiple portrait clusters can be grouped together.

[0060] S203 , recording the to-be-processed images corresponding to the multiple portrait clusters as portrait cluster sets corresponding to the portrait clusters, respectively, to obtain multiple portrait cluster sets containing the same portrait.

[0061] It can be understood that a portrait cluster corresponds to a portrait cluster set, and the images to be processed corresponding to all facial feature maps covered by a portrait cluster are determined as the portrait cluster set corresponding to the portrait cluster, that is, the portrait cluster set contains all the images to be processed corresponding to the portrait clusters containing the same portrait.

[0062] The present invention achieves the extraction of facial attribute features from each image to be processed to obtain a facial feature map of each image to be processed; clustering is performed on all facial feature maps to obtain multiple portrait clusters; and the images to be processed corresponding to the multiple portrait clusters are respectively recorded as portrait cluster sets corresponding to the portrait clusters to obtain multiple portrait cluster sets containing the same portrait. In this way, portrait clusters of different portraits can be automatically distinguished by extracting facial attribute features and applying clustering processing, and all images to be processed corresponding to different portrait clusters are merged into portrait cluster sets corresponding to different portraits, and portrait cluster sets of different portraits can be automatically clustered and classified, reducing the process and cost of manual recognition and classification, and improving the accuracy and efficiency of the classification of portrait cluster sets of different portraits.

[0063] S30, performing static portrait analysis on each portrait cluster set to obtain analysis results corresponding to each portrait cluster set.

[0064] Understandably, the processing process of static portrait analysis can be set according to demand. For example, the processing process of static portrait analysis can be to compare each pixel of each portrait image in each portrait cluster set with each other (that is, the comparison between the pixel values ​​of each pixel in the two portrait images). By analyzing the comparison results between all two portrait images, the mean and variance of these comparison results can be calculated. If the portrait in the portrait image is a static portrait, the difference between all portrait images in the corresponding portrait cluster set is extremely small, so that the comparison result is close to zero, and the variance is also close to zero. In this way, if it is consistent with the mean being less than the preset mean, and the variance is very small and close to zero, the analysis result of the portrait cluster set is determined to be a mis-collected, that is, the portrait cluster set is determined to be a mis-collected image. If it is not consistent with the mean being less than the preset mean, or the variance is very small and close to zero, the analysis result of the portrait cluster set is determined to be retained, indicating that the portrait in the portrait cluster set does not belong to a static portrait, but is a real portrait, that is, it is determined to be a normal image.

[0065] Optionally, the processing process of static portrait analysis can also be as follows: first, perform image quality identification on each to-be-processed image in a portrait cluster set to obtain a quality score corresponding to each to-be-processed image; based on the quality score corresponding to each to-be-processed image in the portrait cluster set, sort all to-be-processed images in the portrait cluster set to obtain a corresponding comparison cluster set; secondly, in the comparison cluster set, record the to-be-processed image with a quality score in the median as a median image, perform feature comparison between the median image and each static portrait image in the static image library, and obtain a comparison cluster set of the median image and each static portrait image; Multiple first feature similarity comparison results between the dynamic portrait images are obtained, and the maximum feature similarity comparison result among all the first feature similarity comparison results is obtained; finally, the corresponding analysis result is determined according to the maximum feature similarity comparison result, that is, if the maximum feature similarity comparison result is greater than or equal to the preset first threshold value, the portrait cluster set is determined to be a missampled image, and if the maximum feature similarity comparison result is less than the preset first threshold value, the portrait cluster set is determined to be a normal image, so that all portrait cluster sets can be traversed and the analysis result corresponding to each portrait cluster set can be obtained.

[0066] Optionally, the processing process of static portrait analysis can also be as follows: first, perform feature comparison on each image to be processed in a portrait cluster set with each static portrait image to obtain multiple second feature similarity comparison results between each image to be processed and each static portrait image; second, obtain the maximum feature similarity comparison result and the average feature similarity comparison result among all the second feature similarity comparison results corresponding to the same static portrait image; finally, determine the analysis result corresponding to the portrait cluster set according to the maximum feature similarity comparison result and the average feature similarity comparison result, that is, if the maximum feature similarity comparison result is greater than a preset second threshold, and the average feature similarity comparison result is greater than a preset third threshold, then the portrait cluster set is determined to be a mis-sampled image; if the maximum feature similarity comparison result is less than or equal to the preset second threshold, or the average feature similarity comparison result is less than or equal to the preset third threshold, then the portrait cluster set is determined to be a normal image, so that all portrait cluster sets can be traversed to obtain the analysis result corresponding to each portrait cluster set.

[0067] In this way, different analysis results are obtained through different static portrait analyses. By comparing the analysis results of different static portrait analyses, a static portrait analysis method with better analysis effect can be flexibly selected to accurately analyze a static image.

[0068] In one embodiment, static portrait analysis is performed on each portrait cluster set to obtain analysis results corresponding to each portrait cluster set, including:

[0069] For each portrait cluster set, image quality identification is performed on each image to be processed in the portrait cluster set to obtain a quality score corresponding to each image to be processed in the portrait cluster set.

[0070] Understandably, image quality recognition is performed on each image to be processed in a portrait cluster, and image quality recognition is recognition through an image quality detection model. Image quality detection includes a texture detection network, a gradient detection network, and a clarity detection network. The image quality detection model extracts texture features, gradient features, and clarity features from the input portrait image, that is, the texture features in the image to be processed are extracted through the texture detection network in the image quality detection model. The texture features are the features of the texture of the pixel color. The extracted texture features are softmax classified to classify the image quality results in the texture dimension, that is, the more regular the texture features are and the more hierarchical the features are, the better the shooting quality of the portrait in the image. The gradient features in the image to be processed are extracted by the gradient detection network in the image quality detection model. The gradient features are the features of the gradient transformation (contrast change) of the pixels in the image. The extracted gradient features are classified by softmax, and the image quality results in the gradient dimension are classified. That is, the smoother the gradient change of the gradient feature is in the histogram corresponding to the image, the better the shooting quality of the portrait in the image is. The clarity features in the image to be processed are extracted by the clarity detection network in the image quality detection model. The clarity features are the features of the edge clarity of the pixels in the image. The extracted clarity features are classified by softmax, and the image quality results in the clarity dimension are classified. That is, the more obvious the edge distinction reflected by the clarity features is, the better the shooting quality of the portrait in the image is. The image quality results in the texture dimension, the image quality results in the gradient dimension, and the image quality results in the clarity dimension are processed by the fully connected layer to obtain the quality scores of each image to be processed. The quality score can reflect the shooting quality of the portrait image. The quality score can be a score from 0 to 100, and the quality of the portrait in the image is measured in the range of 0 to 100.

[0071] Based on the quality score corresponding to each to-be-processed image in the portrait cluster set, all to-be-processed images in the portrait cluster set are sorted to obtain a comparison cluster set corresponding to the portrait cluster set.

[0072] Understandably, according to the quality scores of all the images to be processed in a portrait cluster set, all the images to be processed in the portrait cluster set are sorted in descending order or ascending order, so as to obtain the comparison cluster set corresponding to the portrait cluster set.

[0073] Among the to-be-processed images in the comparison cluster set, the to-be-processed image with a quality score at the median is recorded as the median image of the comparison cluster set.

[0074] It can be understood that the median, also known as the median, is the number in the middle position of the comparison cluster set arranged in order. It can also be understood as half of the total number of comparison cluster sets, dividing the comparison cluster set into two equal parts. If the total number of comparison cluster sets is an even number, the two middle numbers are taken as the median, and the image to be processed in the median of the comparison cluster set is recorded as the median image, that is, when the comparison cluster set is an odd number, the image to be processed corresponding to the middle number is taken as the median image, and when the comparison cluster set is an even number, the images to be processed corresponding to the two middle numbers are taken as the median image.

[0075] A feature comparison is performed between the median image of the comparison cluster set and each static portrait image in the static image library to obtain multiple first feature similarity comparison results between the median image of the comparison cluster set and each static portrait image, and the maximum feature similarity comparison result among the multiple first feature similarity comparison results is obtained.

[0076] Understandably, the feature comparison between the median image and the static portrait image can be achieved through a trained feature comparison model, through which the static features and character features of the two images are respectively extracted, the static features between the median image and the static portrait image are compared to obtain the static comparison result, and the character features between the median image and the static portrait image are compared to obtain the character comparison result, and the static comparison result and the character comparison result are combined, for example, the static comparison result and the character comparison result are weightedly summed to obtain the feature comparison process of the first feature similarity comparison result between the median image and the static portrait image, thereby obtaining multiple first feature similarity comparison results between the median image of the comparison cluster set and each static portrait image, and obtaining the maximum feature similarity comparison result among all the first feature similarity comparison results corresponding to the same portrait cluster set.

[0077] Among them, static features are static and unchanging features in static portrait images collected historically. The learning of static features is a feature comparison model that learns the common features in all static portrait images. Character features are features of characters contained in static portrait images collected historically. The learning of character features is a feature comparison model that learns the common features of postures or expressions of characters in all static portrait images.

[0078] If the maximum feature similarity comparison result is greater than or equal to a preset first threshold, the portrait cluster set is determined to be a mis-sampled image.

[0079] It is understandable that the first threshold can be set according to demand. If the maximum feature similarity comparison result is greater than or equal to the first threshold, all the images to be processed in the portrait cluster are determined to be mis-collected images.

[0080] If the maximum feature similarity comparison result is less than a preset first threshold, the portrait cluster set is determined to be a normal image.

[0081] It can be understood that if the maximum feature similarity comparison result is less than the first threshold, all the to-be-processed images in the portrait clustering set are determined to be normal images.

[0082] The present invention realizes that, for each portrait cluster set, image quality identification is performed on each to-be-processed image in the portrait cluster set to obtain a quality score corresponding to each to-be-processed image in the portrait cluster set; based on the quality score corresponding to each to-be-processed image in the portrait cluster set, all to-be-processed images in the portrait cluster set are sorted to obtain a comparison cluster set corresponding to the portrait cluster set; among the to-be-processed images in the comparison cluster set, the to-be-processed image with a quality score in the median is recorded as the median image of the comparison cluster set; feature comparison is performed on the median image of the comparison cluster set with each static portrait image in the static image library to obtain multiple first feature similarity comparison results between the median image of the comparison cluster set and each static portrait image, and the similarity comparison results of the multiple first feature similarity comparison results are obtained. The maximum feature similarity comparison result is obtained; if the maximum feature similarity comparison result is greater than or equal to the preset first threshold value, the portrait cluster set is determined to be a mis-collected image; if the maximum feature similarity comparison result is less than the preset first threshold value, the portrait cluster set is determined to be a normal image. In this way, the quality score of each portrait image can be automatically identified through image quality recognition, and each portrait cluster set is sorted according to the quality score, and the median image is extracted. Each median image is automatically compared with each static portrait image, and the analysis result of each portrait cluster set is determined according to the maximum comparison result. It is automatically analyzed which portrait cluster sets are mis-collected, which reduces the manual recognition process and greatly improves the accuracy and correctness of image processing for distinguishing normal images from mis-collected images.

[0083] In one embodiment, the image processing method further includes:

[0084] The median image of the portrait cluster set and the static portrait image corresponding to the maximum feature similarity comparison result are compared for image quality to obtain a quality comparison result.

[0085] It can be understood that image quality comparison is a process of detecting the quality score of the median image through an image quality detection model, and detecting the quality score of the static portrait image corresponding to the maximum feature similarity comparison result through an image quality detection model, and comparing the quality scores of the two, so as to obtain a quality comparison result. The quality comparison result reflects which one has better image quality between the missampled median image and the static portrait image corresponding to the maximum feature similarity comparison result. The quality comparison result is the ratio between the missampled median image and the static portrait image corresponding to the maximum feature similarity comparison result.

[0086] If the quality comparison result is greater than a preset quality threshold, the static portrait image corresponding to the maximum feature similarity comparison result is updated in the static image library according to the median image of the portrait cluster set.

[0087] It can be understood that if the quality comparison result is greater than the preset quality threshold, preferably, the preset quality threshold is 1, the corresponding static portrait image will be updated, that is, the median image will cover the current static portrait image, and the median image will be stored in the corresponding static portrait image position in the static image library.

[0088] In this way, the static gallery can be automatically updated, so that the static portrait images in the static gallery can be updated with better quality static portrait images in real time without manual addition, thereby improving the real-time performance and high quality of the static gallery.

[0089] In one embodiment, static portrait analysis is performed on each portrait cluster set to obtain analysis results corresponding to each portrait cluster set, including:

[0090] For each portrait cluster set, feature comparison is performed between each image to be processed in the portrait cluster set and each static portrait image to obtain a plurality of second feature similarity comparison results between each image to be processed and each static portrait image.

[0091] Understandably, feature comparison is performed between each image to be processed in a portrait cluster and each static portrait image. The feature comparison between the image to be processed and the static portrait image is as follows: first reduce the size to a preset size, simplify the color (grayscale processing) of the reduced-size image, and then use the cosine similarity algorithm to compare the average values ​​of the two simplified colors to obtain multiple second feature similarity comparison results between the image to be processed and each static portrait image, and traverse all images to be processed to obtain multiple second feature similarity comparison results between each image to be processed and each static portrait image.

[0092] The cosine similarity algorithm is an algorithm that calculates the cosine value between two numerical values ​​or vector values ​​as the similarity.

[0093] A maximum feature similarity comparison result and an average feature similarity comparison result among all second feature similarity comparison results corresponding to the same static portrait image are obtained.

[0094] It can be understood that by calculating the average value of all second feature similarity comparison results corresponding to the same static portrait image, the average similarity among all second feature similarity comparison results corresponding to the same static portrait image is obtained, and the maximum value of all second feature similarity comparison results corresponding to the same static portrait image is obtained, the maximum feature similarity comparison result among all second feature similarity comparison results corresponding to the same static portrait image can be obtained.

[0095] If the maximum feature similarity comparison result is greater than a preset second threshold, and the average feature similarity comparison result is greater than a preset third threshold, the portrait cluster set is determined to be a mis-captured image.

[0096] It is understandable that if the maximum feature similarity comparison result is greater than the preset second threshold, and the average feature similarity comparison result is greater than the preset third threshold, all the to-be-processed images in the portrait cluster are determined to be mis-captured images.

[0097] If the maximum feature similarity comparison result is less than or equal to the preset second threshold, or the average feature similarity comparison result is less than or equal to the preset third threshold, it is determined that the portrait cluster set is a normal image.

[0098] It is understandable that if the maximum feature similarity comparison result is less than or equal to the preset second threshold, or the average feature similarity comparison result is less than or equal to the preset third threshold, all the images to be processed in the portrait cluster are determined to be normal images.

[0099] The present invention realizes that, for each portrait cluster set, each image to be processed in the portrait cluster set is compared with each static portrait image to obtain multiple second feature similarity comparison results of each image to be processed and each static portrait image; the maximum feature similarity comparison result and the average feature similarity comparison result among all the second feature similarity comparison results corresponding to the same static portrait image are obtained; if the maximum feature similarity comparison result is greater than a preset second threshold value, and the average feature similarity comparison result is greater than a preset third threshold value, the portrait cluster set is determined to be a mis-picked image; if the maximum feature similarity comparison result is less than or equal to the preset second threshold value, or the average feature similarity comparison result is less than or equal to the preset third threshold value, the portrait cluster set is determined to be a normal image, so that the maximum similarity and the average similarity are automatically obtained by calculating the similarity between each image to be processed and each static portrait image, and which portrait cluster sets are mis-picked are automatically analyzed by the maximum similarity and the average similarity, thereby reducing the process of manual recognition and greatly improving the accuracy and correctness of mis-picked portraits.

[0100] S40, based on the analysis results corresponding to each portrait cluster set, performing mis-collection filtering on each portrait cluster set, and generating image collection results for each portrait cluster set.

[0101] It can be understood that false positive filtering is the process of filtering the portrait cluster sets corresponding to the analysis results confirmed as false positive images from all portrait cluster sets, that is, all the images to be processed in the portrait cluster sets corresponding to the analysis results confirmed as false positive images are removed or discarded, and each removed or discarded portrait cluster set is generated as an image acquisition result of a static portrait, all the images to be processed in the portrait cluster sets corresponding to the analysis results confirmed as normal images are retained, and each retained portrait cluster set is generated as an image acquisition result of a real portrait, and at the same time, each retained portrait cluster set after false positive filtering is subjected to high-quality screening, and high-quality images of each portrait cluster set are screened out, the high-quality images are displayed on the front-end visualization interface, and each high-quality image is associated with the corresponding portrait cluster set and stored for subsequent retrieval of the portrait cluster set or / and the high-quality image of the real portrait.

[0102] The present invention achieves the following steps: acquiring a plurality of images to be processed; performing portrait clustering processing on the plurality of images to be processed to obtain a plurality of portrait cluster sets containing the same portrait; performing static portrait analysis on each portrait cluster set to obtain analysis results corresponding to each portrait cluster set; performing mis-collection filtering on each portrait cluster set based on the analysis results corresponding to each portrait cluster set to generate image acquisition results for each portrait cluster set. Therefore, the present invention automatically acquires images to be processed containing portraits, and automatically analyzes mis-collected portrait cluster sets through portrait clustering processing and static portrait analysis, and generates image acquisition results for each portrait cluster set after mis-collection filtering, thereby avoiding interference from static portraits, reducing subsequent image processing operations on static portrait images, and improving the accuracy and correctness of image processing of real portraits.

[0103] In one embodiment, in step S40, the image acquisition results for each portrait cluster set are generated, including:

[0104] The portrait clusters after the mis-collection filtering are subjected to high-quality screening to select high-quality images of each portrait cluster.

[0105] Understandably, high-quality screening is the process of performing face recognition on the images to be processed in each portrait cluster set retained after the false positive filtering, that is, performing face recognition on the images to be processed of the normal images, identifying the face area, and screening the images based on the face features and clarity features in the identified face area in the portrait image. Facial features are features that all have on the face, eyes, nose, mouth, and eyebrows, and clarity features are features of the edge clarity of pixels in the image. A facial feature map is output based on the extracted facial features, and a clarity feature map is output based on the extracted clarity features. Combined with the facial feature map and the clarity feature map, dimensionality reduction processing is performed to obtain the corresponding quality score, and the image to be processed corresponding to the highest quality score among all the images to be processed is screened out, and recorded as a high-quality image of the portrait cluster set.

[0106] Each high-quality image is displayed on a front-end visualization interface, and the high-quality image of each portrait cluster set and the portrait cluster set corresponding to the high-quality image of each portrait cluster set are associated and stored.

[0107] Understandably, the front-end visualization interface allows users or customers to view the image acquisition results of the portrait cluster set through the terminal they hold, package each high-quality image, and make the packaged high-quality image displayable on the front-end visualization interface to view the high-quality image. At the same time, the high-quality images of the portrait cluster set and the portrait cluster sets corresponding to the high-quality images of the portrait cluster set are associated and stored for subsequent calls.

[0108] The present invention achieves high-quality screening of each portrait cluster set after mis-collection filtering to screen out high-quality images of each portrait cluster set; each high-quality image is displayed on a front-end visualization interface, and the high-quality image of each portrait cluster set and the portrait cluster set corresponding to the high-quality image of each portrait cluster set are associated and stored. In this way, through high-quality screening, the high-quality image can be displayed on the front-end visualization interface, and the high-quality image can be associated with the corresponding portrait cluster set, so as to subsequently call the portrait cluster set of the real portrait and subsequently process the image of the real portrait.

[0109] In one embodiment, after the image acquisition results for each portrait cluster set are generated, the following steps are included:

[0110] S50, performing face recognition on the high-quality images in each portrait cluster set to obtain a recognition result corresponding to each portrait cluster set.

[0111] Understandably, the process of performing face recognition on high-quality images in each portrait cluster can be achieved through a trained face recognition model, and the high-quality images are input into the face recognition model. The network structure of the face recognition model is the network structure of YOLO, because the use of the network structure of YOLO can quickly identify the regional results of the face. The training process of the face recognition model is to input a face sample containing a face label, extract face features (convolution) for the face sample, obtain a small-scale feature map based on the extracted face features, upsample and tensor concatenate the small-scale feature map to obtain a medium-scale feature map, and then upsample and tensor concatenate the medium-scale feature map to obtain a large-scale feature map, and then perform multi-scale detection on the small-scale, medium-scale and large-scale feature maps to predict a sample recognition result of a face with face features, calculate the loss value between the sample recognition result and the face label, and iteratively update the parameters of the face recognition model according to the loss value until the loss is reached. The loss value reaches the convergence condition, thereby stopping the training process. Among them, the small scale, medium scale and large scale can be 1 / 32 scale, 1 / 16 scale and 1 / 8 scale relative to the size of the image to be recognized, respectively. Feature maps of different scales can detect more fine-grained facial features. Upsampling is the reverse propagation of convolution, that is, the transposition of deconvolution or forward propagation. Tensor splicing is the process of splicing the upsampling results of the intermediate layer and the adjacent next layer in the convolution process. Facial features are features related to facial parts, such as eyes, face, nose, mouth, eyebrows, etc. that can reflect the outline of a portrait. The recognition process of the face recognition model is to extract facial features from the input image to be recognized through the face recognition model, perform multi-scale detection through small-scale, medium-scale and large-scale feature maps, and predict the recognition result of a face with facial features, so as to determine the recognition result of whose face the face belongs to. The recognition result shows whose face all the images to be processed in the portrait cluster contain.

[0112] In this way, facial feature extraction can be used, and the face recognition model can be used to automatically identify the person in the portrait cluster set.

[0113] It should be understood that the order of execution of the steps in the above embodiment does not necessarily mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present invention.

[0114] In one embodiment, an image processing device is provided, which corresponds one-to-one to the image processing method in the above embodiment. Figure 5 As shown, the image processing device includes an acquisition module 11, a clustering module 12, an analysis module 13 and a filtering module 14. The functional modules are described in detail as follows:

[0115] An acquisition module 11 is used to acquire a plurality of images to be processed;

[0116] A clustering module 12 is used to perform portrait clustering processing on the plurality of images to be processed to obtain a plurality of portrait cluster sets containing the same portrait;

[0117] An analysis module 13, configured to perform static portrait analysis on each of the portrait cluster sets to obtain analysis results corresponding to each of the portrait cluster sets;

[0118] The filtering module 14 is used to perform mis-collection filtering on each of the portrait cluster sets based on the analysis results corresponding to each of the portrait cluster sets, and generate image collection results for each of the portrait cluster sets.

[0119] For the specific definition of the image processing device, please refer to the definition of the image processing method above, which will not be repeated here. Each module in the above image processing device can be implemented in whole or in part by software, hardware and a combination thereof. The above modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0120] In one embodiment, a computer device is provided. The computer device may be a client or a server. The internal structure diagram thereof may be as follows: Figure 6 As shown. The computer device includes a processor, a memory, a network interface and a database connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a readable storage medium and an internal memory. The readable storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the readable storage medium. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, an image processing method is implemented.

[0121] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the image processing method in the above embodiment is implemented.

[0122] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the image processing method in the above embodiment is implemented.

[0123] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided by the present invention may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0124] Those skilled in the art can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0125] The embodiments described above are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included in the protection scope of the present invention.

Claims

1. An image processing method, characterized in that: include: Acquire multiple images to be processed; Performing portrait clustering processing on the plurality of images to be processed to obtain a plurality of portrait cluster sets containing the same portrait; Performing static portrait analysis on each of the portrait cluster sets to obtain analysis results corresponding to each of the portrait cluster sets; Based on the analysis results corresponding to each of the portrait cluster sets, mis-collection filtering is performed on each of the portrait cluster sets to generate image collection results for each of the portrait cluster sets; The performing static portrait analysis on each of the portrait cluster sets to obtain analysis results corresponding to each of the portrait cluster sets includes: For each portrait cluster set, image quality identification is performed on each to-be-processed image in the portrait cluster set to obtain a quality score corresponding to each to-be-processed image in the portrait cluster set; Based on the quality score corresponding to each to-be-processed image in the portrait cluster set, all to-be-processed images in the portrait cluster set are sorted to obtain a comparison cluster set corresponding to the portrait cluster set; Among the to-be-processed images in the comparison cluster set, recording the to-be-processed image whose quality score is at the median as the median image of the comparison cluster set; Performing feature comparison between the median image of the comparison cluster set and each static portrait image in the static image library, obtaining a plurality of first feature similarity comparison results between the median image of the comparison cluster set and each of the static portrait images, and obtaining a maximum feature similarity comparison result among the plurality of first feature similarity comparison results; If the maximum feature similarity comparison result is greater than or equal to a preset first threshold, determining that the portrait cluster set is a mis-collected image; If the maximum feature similarity comparison result is less than a preset first threshold, it is determined that the portrait cluster set is a normal image.

2. The image processing method according to claim 1, characterized in that: The step of acquiring a plurality of images to be processed comprises: Acquire multiple candidate images; Performing single person detection on each of the candidate images to obtain a detection result of each of the candidate images; The candidate image corresponding to the detection result containing the human portrait is determined as the image to be processed.

3. The image processing method according to claim 1, characterized in that: The performing portrait clustering processing on the plurality of images to be processed to obtain a plurality of portrait cluster sets containing the same portrait comprises: Extracting facial attribute features from each of the images to be processed to obtain facial feature maps of each of the images to be processed; Performing clustering processing on all the facial feature graphs to obtain multiple portrait clusters; The images to be processed corresponding to the multiple portrait clusters are respectively recorded as portrait cluster sets corresponding to the portrait clusters, so as to obtain multiple portrait cluster sets containing the same portrait.

4. The image processing method according to claim 1, wherein: The image processing method further includes: Performing image quality comparison between the median image of the portrait cluster set and the static portrait image corresponding to the maximum feature similarity comparison result to obtain a quality comparison result; If the quality comparison result is greater than a preset quality threshold, then in the static image library, the static portrait image corresponding to the maximum feature similarity comparison result is updated according to the median image of the portrait cluster set.

5. The image processing method according to claim 1, wherein: The performing static portrait analysis on each of the portrait cluster sets to obtain analysis results corresponding to each of the portrait cluster sets includes: For each portrait cluster set, feature comparison is performed between each of the images to be processed in the portrait cluster set and each of the static portrait images to obtain a plurality of second feature similarity comparison results between each of the images to be processed and each of the static portrait images; Obtaining a maximum feature similarity comparison result and an average feature similarity comparison result among all the second feature similarity comparison results corresponding to the same static portrait image; If the maximum feature similarity comparison result is greater than a preset second threshold, and the average feature similarity comparison result is greater than a preset third threshold, determining that the portrait cluster set is a mis-collected image; If the maximum feature similarity comparison result is less than or equal to a preset second threshold, or the average feature similarity comparison result is less than or equal to a preset third threshold, it is determined that the portrait cluster set is a normal image.

6. The image processing method according to claim 1, wherein: The generating of the image acquisition results for each of the portrait cluster sets includes: Performing high-quality screening on each of the portrait cluster sets after the false positives have been filtered out, and screening out high-quality images of each of the portrait cluster sets; Each of the high-quality images is displayed on a front-end visualization interface, and the high-quality images of each of the portrait cluster sets and the portrait cluster sets corresponding to the high-quality images of each of the portrait cluster sets are associated and stored.

7. An image processing device, characterized in that: include: An acquisition module, used for acquiring a plurality of images to be processed; A clustering module, used for performing portrait clustering processing on the plurality of images to be processed to obtain a plurality of portrait cluster sets containing the same portrait; An analysis module, used to perform static portrait analysis on each of the portrait cluster sets to obtain analysis results corresponding to each of the portrait cluster sets; A filtering module, configured to filter mis-collected images of each of the portrait cluster sets based on the analysis results corresponding to each of the portrait cluster sets, and generate image collection results for each of the portrait cluster sets; The performing static portrait analysis on each of the portrait cluster sets to obtain analysis results corresponding to each of the portrait cluster sets includes: For each portrait cluster set, image quality identification is performed on each to-be-processed image in the portrait cluster set to obtain a quality score corresponding to each to-be-processed image in the portrait cluster set; Based on the quality score corresponding to each to-be-processed image in the portrait cluster set, all to-be-processed images in the portrait cluster set are sorted to obtain a comparison cluster set corresponding to the portrait cluster set; Among the to-be-processed images in the comparison cluster set, recording the to-be-processed image whose quality score is at the median as the median image of the comparison cluster set; Performing feature comparison between the median image of the comparison cluster set and each static portrait image in the static image library, obtaining a plurality of first feature similarity comparison results between the median image of the comparison cluster set and each of the static portrait images, and obtaining a maximum feature similarity comparison result among the plurality of first feature similarity comparison results; If the maximum feature similarity comparison result is greater than or equal to a preset first threshold, determining that the portrait cluster set is a mis-collected image; If the maximum feature similarity comparison result is less than a preset first threshold, it is determined that the portrait cluster set is a normal image.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the image processing method according to any one of claims 1 to 6 is implemented.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the image processing method according to any one of claims 1 to 6 is implemented.

Images