Human attribute detection method, device, computer equipment and storage medium

Abstract

The present invention relates to the field of artificial intelligence technology, and discloses a human attribute detection method, device, computer equipment and storage medium, the method comprising: preprocessing an image to be detected to obtain an image to be processed; performing human quality and attribute recognition on the image to be processed through a human attribute detection model to obtain a human quality evaluation result and a human attribute recognition result; performing confidence correction processing on the human attribute recognition result according to the human quality evaluation result to obtain a human recognition correction result; generating a human attribute final result corresponding to the image to be detected according to the human quality evaluation result and the human recognition correction result. Therefore, the present invention realizes automatic correction of human attribute recognition results based on the results of human quality recognition, without the cost of manual correction, improves the accuracy of human attribute recognition, and enhances the efficiency and effectiveness of human attribute recognition.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a human attribute detection method, device, computer equipment and storage medium. Background Art

[0002] With the development of artificial intelligence and the large-scale deployment of high-definition video surveillance equipment, human attribute recognition has good application prospects in video surveillance, smart retail, human weight recognition and other fields. It has attracted the attention of more and more researchers and has become a new research topic in the field of video surveillance systems.

[0003] In the prior art, human attribute recognition is usually performed by extracting the structured attributes of the human body in the image (generally including gender, age group, top type, top color, etc.), thereby identifying the results corresponding to different human attributes. However, since the human body may be of poor quality in the image, such as the human body occupies a small area of ​​the image, body parts are occluded, and the face is occluded, etc., this often leads to low effectiveness and subsequent applicability of the identified results, resulting in low accuracy of the results of identifying human attributes. The identified results must be manually corrected at a later stage, which consumes a lot of manpower and material resources and has low recognition efficiency. Summary of the invention

[0004] The present invention provides a human attribute detection method, device, computer equipment and storage medium, which realize automatic correction of human attribute recognition results in combination with human mass recognition results, improve the accuracy of human attribute recognition, save the cost of manual correction, improve the efficiency and effectiveness of human attribute recognition, and improve the applicability of human attribute recognition results.

[0005] A method for detecting human attributes, comprising:

[0006] Preprocessing the image to be detected to obtain an image to be processed;

[0007] Performing human body quality and attribute recognition on the image to be processed by using a human body attribute detection model to obtain a human body quality evaluation result and a human body attribute recognition result;

[0008] According to the human body quality evaluation result, the human body attribute recognition result is processed for confidence correction to obtain a human body recognition correction result;

[0009] A final result of human attributes corresponding to the image to be detected is generated according to the human quality evaluation result and the human recognition correction result.

[0010] A human attribute detection device, comprising:

[0011] An acquisition module is used to preprocess the image to be detected to obtain an image to be processed;

[0012] The recognition module is used to perform human body quality and attribute recognition on the image to be processed through a human body attribute detection model to obtain a human body quality evaluation result and a human body attribute recognition result;

[0013] A correction module, used for performing confidence correction processing on the human attribute recognition result according to the human quality evaluation result to obtain a human recognition correction result;

[0014] The determination module is used to generate a final result of human attributes corresponding to the image to be detected according to the human quality evaluation result and the human recognition correction result.

[0015] 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 human attribute detection method when executing the computer program.

[0016] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the human attribute detection method are implemented.

[0017] The human attribute detection method, device, computer equipment and storage medium provided by the present invention obtain the image to be processed by preprocessing the image to be detected; perform human quality and attribute recognition on the image to be processed by using a human attribute detection model to obtain a human quality evaluation result and a human attribute recognition result; perform confidence correction processing on the human attribute recognition result according to the human quality evaluation result to obtain a human recognition correction result; generate a human attribute final result corresponding to the image to be detected according to the human quality evaluation result and the human recognition correction result. Therefore, the present invention realizes that the image to be detected can be preprocessed and the preprocessed image can be automatically corrected. The processed image is used to perform human body quality and attribute recognition, and the human body quality evaluation results and human body attribute recognition results of the human body in the image are identified. The human body attribute recognition results are automatically corrected in confidence according to the human body quality evaluation results, and finally an accurate human attribute final result with a high confidence level is generated. In this way, the human body quality and human body attributes of the human body in the image can be recognized, and the human body attribute recognition results can be automatically corrected in combination with the results of human body quality recognition, without the cost of manual correction, thereby greatly saving labor costs, improving the accuracy of human body attribute recognition, and enhancing the efficiency and effectiveness of human body attribute recognition, as well as the applicability of human body attribute recognition results. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] 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.

[0019] Figure 1 is a schematic diagram of an application environment of a human attribute detection method according to an embodiment of the present invention;

[0020] Figure 2 is a flow chart of a method for detecting human attributes in one embodiment of the present invention;

[0021] Figure 3 is a flow chart of step S10 of a method for detecting human attributes in one embodiment of the present invention;

[0022] Figure 4 is a flow chart of step S20 of a method for detecting human attributes in one embodiment of the present invention;

[0023] Figure 5 is a flow chart of step S203 of the human attribute detection method in one embodiment of the present invention;

[0024] Figure 6 is a principle block diagram of a human attribute detection device in one embodiment of the present invention;

[0025] Figure 7 is a schematic diagram of a computer device in one embodiment of the present invention. DETAILED DESCRIPTION

[0026] 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.

[0027] The human attribute detection method provided by the present invention can be applied in 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.

[0028] In one embodiment, if Figure 2As shown, a method for detecting human attributes is provided, and its technical solution mainly includes the following steps S10-S40:

[0029] S10, preprocessing the image to be detected to obtain an image to be processed.

[0030] It can be understood that the image to be detected is an image containing a single human body and the categories of human body related attributes need to be identified, wherein the categories of human body related attributes can be age, gender, human body orientation, hair color, hair length, type of headwear, whether wearing glasses, whether carrying a backpack, whether carrying a shoulder bag, whether strolling, whether holding a child, whether carrying a handbag, whether wearing a coat, whether riding, whether carrying a trolley case, whether answering a call, top color, top texture, top type, bottom color, bottom texture, bottom type, etc., which can reflect the categories of structured attributes of the human body. The method of acquiring the image to be detected can be to intercept or extract an image of a single human body from an image captured by an acquisition device or a surveillance camera, and the preprocessing process can be performed as needed. Setting, for example, preprocessing can be image enhancement processing on the input image to enhance the human body information in the image to facilitate the subsequent identification of human body quality and attributes, or image segmentation processing can be performed on the input image to segment channel images with different channel information. The channel information carried by different channel images can be beneficial to the identification of human body attributes. Only valid information is required to be identified to avoid interference from other channel information on its identification. Image enhancement processing and image segmentation processing can also be combined to process the input image together, so that the image to be processed can be output. The image to be processed is the image to be detected after the preprocessing, that is, the image to be processed is an image in which information that is beneficial to the identification of human body quality and attributes is extracted from the image to be detected.

[0031] In one embodiment, before the image to be detected is preprocessed, the process includes:

[0032] An input image is acquired and input into a single body recognition model; the input image contains at least one human body.

[0033] It can be understood that the input image is an image collected by an acquisition device or a surveillance camera, the input image can be an image frame containing a human body extracted from a surveillance video, the input image contains at least one human body, and the method of obtaining the image to be input can be set according to needs, such as real-time acquisition or extraction from a historical video, etc. The single-body recognition model is a trained target detection model for identifying the area of ​​a single human body in the input image, and the network structure of the single-body recognition model can be set according to needs, such as the network structure of the single-body recognition model can be Faster The network structures of R-CNN, SSD and YOLO are preferred. Preferably, the network structure of the monomer recognition model is the network structure of CenterNet. The training process of the monomer recognition model is to input a sample image for training, the sample image contains at least one human body, and the sample image is associated with a region annotation of the human body, generate 1K to 2K candidate regions for the sample image, and then for each candidate region, use the convolutional layer of the network structure of the model to extract monomer features, send the extracted monomer features to the human classifier, determine whether the candidate region contains a human body, use a regressor to finely correct the position of the candidate frame, and finally identify the region containing a complete human body in the sample image, calculate the loss value between the region and the region annotation, iteratively update the parameters of the monomer recognition model according to the loss value, until the loss value reaches the convergence condition, thereby stopping the training process.

[0034] The input image is subjected to monomer feature extraction by the monomer recognition model, and monomer recognition is performed according to the extracted monomer features to recognize at least one monomer region containing a human body.

[0035] It can be understood that the monomer feature is a feature of a single human body, such as head, hair, hand, face, torso, clothes, legs, feet, etc., which can reflect the characteristics of a human body. The input image is convolved with the monomer feature through the monomer recognition model, and the monomer feature in the input image is extracted layer by layer using a downsampling method, and monomer recognition is performed based on the extracted monomer feature. The monomer recognition process is a process of identifying a rectangular area with the monomer feature, so that each human body can be identified from the input image, and the monomer area corresponding to each human body is framed. The monomer area reflects the area framed by a human body in the input image, and will be framed regardless of whether the human body in the input image is truncated or blocked.

[0036] An image corresponding to each of the monomer regions is extracted from the input image to obtain an image to be detected corresponding to each of the monomer regions one by one.

[0037] It can be understood that the corresponding image is obtained by cutting out or copying the range framed by each monomer area from the input image to obtain the image to be detected corresponding to each monomer area, which means that the image to be detected contains a single human body.

[0038] The present invention achieves the following steps: acquiring an input image and inputting the input image into a monomer recognition model; the input image contains at least one human body; performing monomer feature extraction on the input image through the monomer recognition model, performing monomer recognition based on the extracted monomer features, and identifying at least one monomer region containing a human body; extracting images corresponding to each of the monomer regions from the input image, and obtaining images to be detected corresponding to each of the monomer regions one by one. In this way, the monomer feature extraction of the input image can be performed through the monomer recognition model, and the image to be detected containing a single human body in the input image can be automatically extracted without manual extraction, thereby saving the cost of manual extraction from massive images one by one, and improving the efficiency of extracting images to be detected.

[0039] In one embodiment, if Figure 3 As shown, in the step S10, that is, preprocessing the image to be detected to obtain the image to be processed, includes:

[0040] S101, performing image enhancement processing on the image to be detected to obtain an enhanced image.

[0041] It is understandable that the image enhancement processing method can be set according to needs. For example, image enhancement may include methods of removing noise, contrast enhancement, color enhancement, etc., or a processing method of performing image completion on the image to be detected before image enhancement processing. The image completion is a method of supplementing the truncated or occluded parts in the image to be detected so that the image to be detected can contain a complete human body. The enhanced image is the image to be detected after the image enhancement processing, and the enhanced image can further highlight the human body information in the image to be detected.

[0042] Among them, the noise removal is a process of transforming each pixel in the input image using a Fourier transform algorithm, the contrast enhancement is to enhance the contrast between black and white to make the boundary or edge in the image more obvious, the color enhancement is to convert the image of each pixel in the image into a color space of another spatial domain, subtract the minimum value of all pixels in the spatial domain from the pixel value of each pixel, and then convert it into the space of the original domain, and the grayscale processing is a method of calculating the pixel value of each pixel in the input image using a grayscale algorithm to obtain an image.

[0043] In one embodiment, performing image enhancement processing on the image to be detected to obtain an enhanced image includes:

[0044] The human body truncation detection model is used to perform limb truncation recognition on the image to be detected to obtain a truncation recognition result.

[0045] It can be understood that the human body truncation detection model is a trained neural network model for identifying whether the human body is truncated or occluded in the input image, and the process of limb stage recognition is to perform limb key point recognition and head recognition on the image to be detected, wherein the process of limb key point recognition is to extract limb features of the input image to be detected through the limb key point detection network, locate the limb points according to the extracted limb features, locate the position and area of ​​the limb points of the human body, and classify the limb types according to the area of ​​the located limb points, and classify each located limb point into head, left / right hand, left / right arm, torso, left / right leg and left / right foot, so that the position of each limb point can be classified. The recognition result of the corresponding limb key points is obtained, where the limb features are features of a human head, hands, arms, torso, legs and feet, etc. The head recognition process is to extract hair features and face features of the input image to be detected in parallel through the head detection network, locate the hair area according to the extracted hair features, classify the located hair area, and judge whether the area is hair, and locate the face area according to the extracted face features, classify the located face area, and judge whether the area is a face, and finally identify the hair area determined to be hair and the face area determined to be a face, where the hair features are features related to human hair, and the face features are features related to the face, such as eyes, nose, mouth, eyebrows, etc. The limb truncation recognition includes two types of truncation recognition. One type of truncation recognition is to compare the recognition results of the limb key points of the head, left hand, right hand, left arm, right arm, torso, left leg, right leg, left foot, and right foot one by one, and judge whether there is inconsistency or missing in the comparison results, and confirm the missing of any one or more limb points according to the comparison results, so as to determine the limb key point truncation result. The other type of truncation recognition includes judging whether the hair area or the face area is missing, and determining whether the hair or the face is partially truncated according to the proportion of the hair area and the face area, and confirming the head by combining the results of whether the hair area or the face area is missing and whether the hair or the face is partially truncated. The truncation result is obtained according to the limb key point truncation result and the head truncation result, that is, it is determined whether there is a part truncation result in the limb key point truncation result and the head truncation result, so as to use the part truncation result as the truncation recognition result. The truncation recognition result can reflect which limb parts of the human body in the image to be detected are missing. For example, the truncation recognition result is that the truncation part is the right leg and right foot, or the truncation part is the hair, or the truncation part is the left hand and left arm, etc. The truncation recognition result may also be no truncation part or a complete human body. The human body truncation detection model can automatically identify whether there is a truncation part in the image to be detected, which provides an evaluation basis for subsequent human body quality recognition.It is beneficial to the efficiency and accuracy of human body quality recognition. For example, when there is a truncation part in the truncation recognition result, it means that the human body in the captured image is incomplete and the image quality of the human body is poor. It can be used as a factor in the subsequent occlusion score or occlusion result output, making the occlusion score or occlusion result more accurate.

[0046] When a truncation part is detected in the truncation recognition result, a truncation part prediction corresponding to the truncation part is performed on the image to be detected to obtain a predicted area, and image completion is performed on the image to be detected according to the predicted area, and image enhancement processing is performed on the image to be detected after image completion to obtain the enhanced image.

[0047] Understandably, when the truncation recognition result is obtained, it is determined whether the truncation recognition result has a truncation part, that is, it is determined whether the truncation part includes the head, left / right hand, left / right arm, torso, left / right leg, left / right foot, hair and face. If any truncation part exists, the symmetric mirror method of the human body is used to predict the truncation part corresponding to the truncation part for the image to be detected.

[0048] Among them, the process of the prediction method of the symmetrical mirror image of the human body is: first, the part symmetrical to the truncated part is identified according to the truncated part, that is, each part has a corresponding mirror-symmetrical part, and according to the correspondence table, the mirror-symmetrical part corresponding to the truncated part can be determined; secondly, the area range of the part marked as mirror-symmetrical is retrieved, and mirror translation is performed according to the area range to obtain the mirror-symmetrical area, that is, the area range involved in the mirror-symmetrical part is mirror-translated to the corresponding symmetrical area according to the center line of the identified torso part as the central axis; finally, only the content of the image to be detected is retained in the intersection part of the symmetrical area and the image to be detected, and the symmetrical area outside the intersection part is determined as the prediction area. Among them, the symmetrical area is a rectangular area, for example, a face and hair can be in a mirror symmetrical relationship. If the cut-off part is the hair, it means that the hair is missing. The area where the hair is cut off can be mirrored according to the identified face area to obtain the predicted area. For example, if the cut-off part is the left leg and the left foot, the human body symmetrical mirroring method is used. The left leg and the right leg are in a mirror symmetrical relationship. The left foot and the right foot are predicted to have the area of ​​the left leg and the left foot according to the mirror image of the right leg and the right foot. Or if the cut-off part is the right shoulder and the right hand, the human body symmetrical mirroring method is used to predict the area of ​​the right shoulder and the right hand according to the mirror image of the left shoulder and the left hand, and the predicted area is recorded as the predicted area. In this way, the image completion is beneficial to the subsequent recognition of human attributes. For example, the image completion of the cut-off left leg and left foot of the human body can be beneficial to the recognition of the movement direction or walking direction of the human body, or the image completion of the cut-off right hand and right shoulder of the human body can be beneficial to the recognition of the movement posture of the human body.

[0049] Among them, the image completion can be understood as supplementing the predicted area from the image edge of the truncated part of the image to be detected to the periphery of the image, and filling each pixel in the predicted area with a preset pixel value of the identifier. The identifier can reflect that the pixel value is the pixel value of image completion. The preset pixel value can be set according to needs. For example, the preset pixel value can be "255" or "0", and even the preset pixel value can be the average of all pixel values ​​in the area corresponding to the part symmetrical to the truncated part. The image to be detected after image completion is used as an extended image. The size of the extended image is expanded according to the image to be detected, expanded to a size that can cover the predicted area, and the area other than the predicted area and the image to be detected in the extended image is filled with blank pixels, so that the area filled with blank pixels has no effect on the subsequent identification of human body quality and attributes, and the image enhancement processing is performed on the extended image to obtain the enhanced image.

[0050] When it is detected that there is no truncation part in the truncation recognition result, image enhancement processing is performed on the image to be detected to obtain the enhanced image.

[0051] It is understandable that when obtaining the truncated recognition result, it is detected that there is no truncated part in the truncated recognition result, that is, the truncated recognition result is a human body without truncated part or a complete human body, and the image enhancement processing is directly performed on the image to be detected to obtain the enhanced image corresponding to the image to be detected.

[0052] The present invention realizes that the limb truncation recognition is performed on the image to be detected by a human body truncation detection model to obtain a truncation recognition result; when a truncation part is detected in the truncation recognition result, a truncation part corresponding to the truncation part is predicted for the image to be detected to obtain a predicted area, and image completion is performed on the image to be detected according to the predicted area, and image enhancement processing is performed on the image to be detected after image completion to obtain the enhanced image; when it is detected that there is no truncation part in the truncation recognition result, image enhancement processing is performed on the image to be detected to obtain the enhanced image, so that the human body truncation detection can be realized. The detection model automatically identifies whether there is a truncated part in the image to be detected. When the truncated part is detected, the image is automatically completed, and then image enhancement processing is performed to obtain an enhanced image for subsequent human attribute recognition. When it is detected that there is no truncated part, the image enhancement processing is directly performed to obtain an enhanced image. Therefore, it can automatically identify whether there is a truncated part in the image to be detected. Based on the recognition result, the subsequent automatic use of image completion can ensure that the human body in the image to be detected is a complete human body, which is convenient for the recognition of human body quality and attributes, and improves the accuracy and efficiency of human body quality, as well as the accuracy and correctness of human body attribute recognition.

[0053] S102, performing image channel splitting and merging on the image to be detected and the enhanced image to obtain the image to be processed.

[0054] It can be understood that the image channel splitting is to perform color segmentation of RGB channels or HSV channels on the input image, the color segmentation of RGB channels is the process of separating three channel images of the three color channels of red (R), green (G) and blue (B) of the image, the color segmentation of HSV channels is the three channel images after the color segmentation of RGB channels, and the three channel images are converted into three converted channel images such as Hue (hue), Saturation (saturation, color purity) and Value (brightness) in the HSV color space, and the merging is to merge the multiple channel images obtained after the image channel splitting of the image to be detected and the enhanced image into a to-be-processed image with a multi-channel image, for example, the image to be detected obtains three channel images after the image channel splitting, and the enhanced image obtains three channel images after the image channel splitting, then the total six channel images are merged to obtain the to-be-processed image with six channel images.

[0055] Wherein, the image to be processed includes at least one channel image.

[0056] The present invention realizes obtaining an enhanced image by performing image enhancement processing on the image to be detected; performing image channel splitting and merging on the image to be detected and the enhanced image to obtain the image to be processed. In this way, image enhancement processing, image channel splitting and merging can be used to automatically obtain an image to be processed that is beneficial to the recognition of human body mass and attributes, thereby improving the efficiency and accuracy of the recognition of human body mass and attributes.

[0057] S20, performing human body quality and attribute recognition on the image to be processed by using a human body attribute detection model to obtain a human body quality evaluation result and a human body attribute recognition result.

[0058] Understandably, the identification of human body quality and attributes of the image to be processed can be achieved through a trained deep learning-based human body attribute detection model for identifying human body quality and attributes. The human body attribute detection model can be a multi-task learning-based neural network model based on deep learning, that is, the identification process of human body quality and the categories of human body attributes is learned through multiple branch tasks, that is, a branch task of human body quality is implemented by a branch for learning whether the human body quality is good or bad, and each branch task of human body attributes is implemented by branches that learn the categories of human body attributes by their own independent tasks. The human body attribute branch tasks can also share a backbone network to be implemented, that is, the human body attribute detection model includes a backbone network, a pooling layer, a human body quality branch network and at least one human body attribute branch network. The backbone network is a trained deep learning-based neural network model. Through the backbone network, shared features of the input image can be learned and shared features can be extracted. The vectors of the extracted shared features can be jointly provided to the human body attributes and human body quality for identification. , reducing the processing load of the processor, and eliminating the need for multi-tasking to extract features in different directions. The pooling layer is a layer for performing pooling operations on the extracted shared feature vectors. The human body quality branch network is a fully connected layer that receives the shared feature map output by the pooling layer and identifies the quality of the input image based on the shared feature map. The human body attribute branch network is a fully connected layer that receives the shared feature map output by the pooling layer and identifies a certain human body attribute category of the human body in the input image based on the shared feature map. The result output by the human body quality branch network is recorded as the human body quality evaluation result, and the results output by all the human body attribute branch networks are summarized to obtain the human body attribute recognition result. The human body quality evaluation result reflects the result of the shooting quality of the human body in the image to be detected. The human body attribute recognition result reflects the result of the set of human body-related attribute categories in the image to be detected. The human body attribute recognition result can be used to obtain the categories of human body-related attributes that need to be paid attention to in the image to be detected, providing a data basis for subsequent screening.

[0059] In one embodiment, if Figure 4As shown, in step S20, that is, the human body quality and attribute recognition is performed on the image to be processed by the human body attribute detection model to obtain a human body quality evaluation result and a human body attribute recognition result, including:

[0060] S201, extracting shared features from the image to be processed through a backbone network to obtain a shared feature vector; the human attribute detection model includes the backbone network, a pooling layer, a human mass branch network and at least one human attribute branch network.

[0061] It can be understood that the backbone network is a trained neural network model based on deep learning. The network structure of the backbone network can be a network structure such as the MobileNet series, ResNet series, etc. Preferably, the network structure of the backbone network is the network structure of ResNet18. The shared features are implicit features learned or convolved in the shared input image between human body mass and various human attributes. The shared features are extracted from each channel image in the image to be processed through the backbone network to obtain the shared feature vector. For example, the input image to be processed is an image with a size of 256 and 128 and contains 6 channels. After the shared features of the backbone network are extracted, a shared feature vector with a shape of (512, 8, 4) is obtained, that is, a feature vector of 8×4 size with 512 channels.

[0062] Among them, the pooling layer is a layer that performs pooling operations on the extracted shared feature vectors, the human body quality branch network is a fully connected layer that receives the shared feature map output by the pooling layer and identifies the quality of the input image based on the shared feature map, and the human body attribute branch network is a fully connected layer that receives the shared feature map output by the pooling layer and identifies a certain human body attribute category of the human body in the input image based on the shared feature map.

[0063] S202, performing global average pooling processing on the shared feature vector through the pooling layer to obtain a shared feature map.

[0064] Understandably, the global average pooling process is also called Global Average Pooling, which means adding all pixel values ​​in the feature vector and averaging them to obtain a numerical value, that is, using this numerical value to represent the corresponding feature map. It can structurally regularize the entire network to prevent overfitting, eliminate the features of the black box operation of the fully connected layer, and directly give each channel the actual human mass or human attribute category meaning, thereby obtaining the shared feature map. For example, after the shared feature vector with a shape of (512, 8, 4) is subjected to global average pooling of the pooling layer, the shape of the feature map obtained is (512, 1, 1).

[0065] S203, performing a human body proportion score on the shared feature map through the human body quality branch network to obtain a proportion result, and performing a human body occlusion score on the shared feature map to obtain an occlusion result, and obtaining the human body quality evaluation result based on the proportion result and the occlusion result.

[0066] Understandably, the shared feature map is scored for the human body proportion through the human body mass branch network, and the human body proportion score can be set according to demand. In one embodiment, the human body proportion score of the shared feature map includes: performing softmax classification processing of the shared feature map at the fully connected layer, and the softmax classification processing is to further transform the score vector output by the linear operation of the classification function by the softmax classifier, that is, the scores of each vector in the shared feature map for each class are converted into a probability distribution through the Softmax function, and the probability of being classified as a human body is used as the human body proportion score; in another embodiment, the shared feature map is classified, and the vector whose probability of having human body features in the shared feature map reaches a preset threshold is upsampled (upsampling) to the vector, and the upsampling is to use the deconvolution method, that is, the inverse process of convolution, to realize transposed convolution (transposed convolution). convolution), obtain an image to be evaluated of the same size as the image to be processed, and mark an area with human body pixels in the image to be evaluated, determine the area as a human body area, extract the facial features and the limb features in the human body area, identify the face area according to the extracted facial features, calculate the area of ​​the image to be evaluated, the area of ​​the face area and the area of ​​the human body area, divide the area of ​​the human body area by the area of ​​the image to be evaluated, calculate the area ratio of the human body area in the image to be evaluated, divide the area of ​​the face area by the area of ​​the image to be evaluated, calculate the area ratio of the face area in the image to be evaluated, and divide the area of ​​the face area by the area of ​​the image to be evaluated, calculate the area ratio of the face area in the image to be evaluated, and calculate the area ratio of the human body area in the image to be evaluated. The area proportion of the estimated image and the area proportion of the face area in the image to be evaluated are averaged or weighted summed to obtain the proportion result, which reflects the result of measuring the quality of the human body in the image to be detected in the proportion dimension, and the proportion result is a percentage for measuring the proportion. The human body occlusion score is performed on the shared feature map through the human body quality branch network. The process of the human body occlusion score is as follows: first, the shared feature map is classified, and a vector with a probability of having human body features and face features reaching a preset threshold is classified from the shared feature map, and the vector is upsampled (upsampling) to obtain a to-be-scored image of the same size as the image to be processed;Secondly, the limb features, facial features and human body features are extracted from the image to be scored, and key limb points, facial areas and human body areas are identified based on the extracted limb features, facial features and human body features. Whether one or more limb points are missing is determined based on the key limb points. In the case of missing limb points, the predicted area is determined using the human body symmetry mirror method. A first proportion score is obtained based on the proportion of the predicted area and the human body area. Whether the hair or face is partially truncated is predicted by calculating the proportion of the facial area and the human body area, and a second proportion score is obtained. The occlusion result is determined by combining the first proportion score and the second proportion score. The process of predicting whether the human body is truncated or occluded in the image to be detected and evaluating the proportion score of the truncation or occlusion in the human body can directly obtain the score. The truncation recognition result and the predicted area are combined, and the area of ​​the predicted area is calculated, without determining whether it is truncated or blocked again, and the area of ​​the truncated or blocked area is calculated, by calculating the area ratio of the predicted area to the face area or the body area, according to the level range corresponding to the area ratio, a high, medium and low score result is determined, and it is determined as the occlusion result, for example: when the truncation part is the head, if the predicted area of ​​the head is truncated is greater than two-thirds of the sum of the face area and the predicted area, it is determined that the truncation or occlusion is serious, and then the score is low, and it is considered that the body quality is poor. When the truncation part is the left shoulder or the left hand, if the predicted area occupies more than one-third of the body area, it is determined that the truncation or occlusion is serious, and then the score is low, and it is considered that the body quality is poor, thereby obtaining the occlusion result of the percentage score of the body quality in the truncation or occlusion dimension. ;

[0067] Among them, according to the good or bad of the proportion result and the good or bad of the occlusion result, as well as their respective percentages, the good or bad of the human body quality evaluation result and the corresponding quality confidence are determined, and the quality confidence measures the confidence level of the good or bad human body quality.

[0068] In one embodiment, if Figure 5 As shown, in step S203, that is, the human body proportion score is performed on the shared feature map through the human body mass branch network to obtain the proportion result, including:

[0069] S2031, upsampling the shared feature map through the human body mass branch network to obtain a map to be evaluated.

[0070] It can be understood that the upsampling process refers to the process of restoring the resolution of the shared feature map to the resolution size of the image to be processed, thereby obtaining the image to be evaluated.

[0071] S2032, performing key limb point recognition, face recognition and body recognition on the image to be evaluated through the human body mass branch network to obtain key limb point coordinate results, face area results and body area results.

[0072] It can be understood that the key limb point recognition is the process of extracting the features of the limb nodes such as the head, hands, arms, torso, legs and feet of the human body in the image to be evaluated, and identifying the limb nodes based on the extracted features. The key limb point recognition is performed on the image to be evaluated, and the key limb point coordinate results can be obtained. The key limb point coordinate results reflect the positions of each key limb node in the image to be evaluated. The face recognition is the process of identifying the area of ​​the face in the image to be evaluated. The face recognition is performed on the image to be evaluated, and the face area result can be obtained. The face area result reflects the area of ​​the face in the image to be evaluated. The human body recognition is the process of identifying the area of ​​the human body in the image to be evaluated. The human body recognition is performed on the image to be evaluated, and the human body area result can be obtained. The human body area result reflects the area of ​​the human body in the image to be evaluated.

[0073] S2033, determining a first proportion value according to the distance between the key limb point coordinate result and the edge of the image to be detected, and determining a second proportion value according to the proportion of the face area result and the body area result in the image to be detected.

[0074] It can be understood that the key limb point coordinate results reflected by the spatial coordinates of the image to be evaluated are correspondingly translated to the spatial coordinates of the image to be detected. The correspondence between the spatial coordinates of the image to be evaluated and the spatial coordinates of the image to be detected is established based on the pixel correspondence between the image to be evaluated and the image to be detected. The minimum vertical distance between the coordinates of each limb node in the key limb point coordinate results and the edge of the image to be detected is calculated. The minimum vertical distance is the minimum vertical distance between a limb node and the four sides of the edge of the image to be detected. According to the minimum vertical distance of each limb node, each limb node is scored accordingly. Finally, the scoring results of all limb nodes are combined to obtain the The first proportion value, for example, calculates the minimum vertical distance of the nodes of the limb head, and determines whether the minimum vertical distance exceeds twice the long side of the face area. If the minimum vertical distance of the nodes of the limb head exceeds twice the long side of the face area, it is determined that the proportion of the head part is small, and the proportion score of the head part is low; calculates the minimum vertical distance of the nodes of the left shoulder of the limb, and determines whether the minimum vertical distance exceeds the short side of the human body area. If the minimum vertical distance of the nodes of the left shoulder of the limb exceeds the short side of the human body area, it is determined that the proportion of the left shoulder part is small, and the proportion score of the left shoulder part is low, and so on. The proportion score of each limb part can be determined, and the first proportion value is obtained by summing and averaging the proportion score results of each limb node.

[0075] Among them, the percentage of the area of ​​the face region in the face region result to the area of ​​the image to be detected, and the percentage of the area of ​​the human body region in the human body region result to the area of ​​the image to be detected are calculated, and the percentage of the area of ​​the face region to the area of ​​the image to be detected and the proportion of the human body region result to the image to be detected are weighted summed to obtain the second proportion value, and the weight of the percentage of the area of ​​the face region to the area of ​​the image to be detected is less than the weight of the proportion of the human body region result to the image to be detected.

[0076] S2034: Weight the first proportion value and the second proportion value to obtain the proportion result.

[0077] It can be understood that the value obtained by multiplying the first proportion value and the corresponding weight is added to the value obtained by multiplying the second proportion value and the corresponding weight to obtain the proportion result, which reflects the result of measuring the quality of the human body based on the proportion dimension of the effective human body part in the image to be detected.

[0078] The present invention realizes upsampling the shared feature map through the human body mass branch network to obtain the map to be evaluated; performing key limb point recognition, face recognition and body recognition on the map to be evaluated through the human body mass branch network to obtain key limb point coordinate results, face area results and body area results; determining a first proportion value according to the distance of the key limb point coordinate results from the edge of the image to be detected, and determining a second proportion value according to the proportion of the face area result and the body area result in the image to be detected; weighting the first proportion value and the second proportion value to obtain the proportion result, so that key limb point recognition, face recognition and body recognition can be used to automatically evaluate the results of human body quality measured in the proportion dimension of the image, thereby improving the data basis for subsequent human body quality recognition and improving the accuracy and reliability of human body quality recognition.

[0079] S204, performing attribute recognition of respective branch tasks on the shared feature graph through each of the human attribute branch networks, obtaining attribute branch results corresponding to each of the human attribute branch networks one by one, and merging all of the attribute branch results into the human attribute recognition result.

[0080] Understandably, the attributes of the respective branch tasks are identified as different human attribute branch networks learning different categories of human attributes, and each of the human attribute branch networks extracts the features of the respective human attribute categories from the shared feature map, and the feature extraction of the respective human attribute categories is trained by various samples containing different human attributes, such as training through the network structure of VGG16, learning the common features of different categories of human attributes, and performing full connection layer classification based on the learned common features of different categories of human attributes, so as to identify the categories of human attributes possessed by each sample, thereby obtaining the trained respective branch tasks for identifying various different human attributes, and the shared The feature maps are respectively input into the human attribute branch networks for different human attribute recognition, and the corresponding features are extracted one by one through the human attribute branch networks for different human attribute recognition, and are classified one by one according to the extracted features, so as to identify the results of the corresponding categories of each different human attribute according to the classification results, so as to identify the attribute branch results corresponding to each of the human attribute branch networks, the attribute branch results reflect the categories recognized by the human attribute branch networks, and include the confidence of the categories, and finally all the attribute branch results are merged into one result, which is recorded as the human attribute recognition result, and the human attribute recognition result includes the categories of each human attribute and the confidence corresponding to the categories of each human attribute.

[0081] The execution order of step S203 and step S204 is not limited here, and the two steps can be executed in serial or in parallel without any particular order.

[0082] The present invention realizes that shared feature extraction is performed on the image to be processed through a backbone network to obtain a shared feature vector; global average pooling is performed on the shared feature vector through the pooling layer to obtain a shared feature map; human body proportion scoring is performed on the shared feature map through the human body quality branch network to obtain a proportion result, and human body occlusion scoring is performed on the shared feature map to obtain an occlusion result, and the human body quality evaluation result is obtained according to the proportion result and the occlusion result; the attribute recognition of the respective branch tasks is performed on the shared feature map through each of the human body attribute branch networks to obtain an attribute branch result corresponding to each of the human body attribute branch networks, and all the attribute branch results are merged into the human body attribute recognition result. In this way, shared feature extraction of the image to be processed is realized by sharing a backbone network, and human body quality evaluation results and human body attribute recognition results are automatically scored based on the shared feature, providing a data basis for subsequent correction of human body attribute recognition.

[0083] S30, performing confidence correction processing on the human attribute recognition result according to the human body quality evaluation result to obtain a human body recognition correction result.

[0084] It can be understood that the confidence correction processing is a processing process for correcting each confidence in the human attribute recognition result according to the good or bad quality of the human quality evaluation result and the quality confidence, that is, first, when the quality score value in the human quality evaluation result is the value corresponding to good quality (1), the quality confidence in the human quality evaluation result is multiplied by each confidence in the human attribute recognition result to obtain the corrected confidence corresponding to each confidence; when the quality score value in the human quality evaluation result is the value corresponding to bad quality (0), the quality confidence in the human quality evaluation result is subtracted from one and the result is multiplied by each confidence in the human attribute recognition result to obtain the corrected confidence corresponding to each confidence; then, the human attribute recognition result is replaced by the corrected confidence corresponding to each confidence, so as to obtain the human recognition corrected result, which is the human attribute recognition result after the confidence correction processing.

[0085] In one embodiment, the human body quality assessment result includes a quality score value and a quality confidence level.

[0086] Understandably, the quality score value may be two scores, including a score value (1) for good quality and a score value (0) for bad quality, and the quality confidence level measures the confidence level of whether the human body quality is good or bad.

[0087] In one embodiment, in step S30, the confidence correction processing is performed on the human attribute recognition result according to the human quality evaluation result to obtain the human recognition correction result, including:

[0088] According to the quality confidence, each attribute branch result in the human attribute recognition result is subjected to confidence correction processing corresponding to the quality score value to obtain a branch correction result corresponding to each attribute branch result one by one.

[0089] It can be understood that each attribute branch result in the human body quality evaluation result and the human body attribute recognition result is input into a confidence correction function, and a branch correction result corresponding to each attribute branch result is calculated by the confidence correction function. The branch correction result includes the category of the human body attribute of the corresponding branch in the human body attribute recognition result and the corresponding correction confidence. The confidence correction function is:

[0090]

[0091] Among them, attr p_revision The corrected confidence in the branch corrected result corresponding to the attribute branch result;

[0092] quality p It is the quality confidence level in the human body quality assessment results;

[0093] quality value is the quality score value in the human body quality evaluation result;

[0094] attr p is the confidence level in the attribute branch result.

[0095] All the branch correction results are aggregated to obtain the human body recognition correction result.

[0096] It can be understood that the correction confidence in the human body recognition correction result can truly reflect the true confidence of the category of human body attributes.

[0097] The present invention achieves the following steps: performing confidence correction processing corresponding to the quality score value on each attribute branch result in the human attribute recognition result according to the quality confidence, thereby obtaining a branch correction result corresponding to each attribute branch result one by one; all the branch correction results are aggregated to obtain the human recognition correction result. In this way, the human attribute recognition result can be automatically corrected in combination with the result of human quality recognition, without the cost of manual correction, thereby greatly saving labor costs, improving the accuracy of human attribute recognition, and enhancing the efficiency and effectiveness of human attribute recognition.

[0098] S40, generating a final result of human attributes corresponding to the image to be detected according to the human quality evaluation result and the human recognition correction result.

[0099] It can be understood that the human body quality evaluation result and the human body recognition correction result are combined to generate a final result of human body attributes corresponding to the image to be detected. The final result of human body attributes can reflect the categories of each human body attribute of the image to be detected and the quality of the human body, which is equivalent to assigning a set of retrieval labels to the image to be detected by the final result of human body attributes. The retrieval label provides retrieval factors for the subsequent retrieval of the image to be retrieved, and the degree of compliance of each factor can be reflected by the correction confidence and quality confidence.

[0100] The present invention achieves the following steps: preprocessing the image to be detected to obtain the image to be processed; identifying the human body quality and attributes of the image to be processed by a human body attribute detection model to obtain a human body quality evaluation result and a human body attribute recognition result; performing confidence correction processing on the human body attribute recognition result according to the human body quality evaluation result to obtain a human body recognition correction result; and generating a human body attribute final result corresponding to the image to be detected according to the human body quality evaluation result and the human body recognition correction result. Therefore, the present invention achieves the following steps: preprocessing the image to be detected, automatically identifying the human body quality and attributes of the preprocessed image, identifying the human body quality evaluation result and the human body attribute recognition result of the human body in the image, and automatically correcting the confidence of the human body attribute recognition result according to the human body quality evaluation result, and finally generating a human body attribute final result with an accurate confidence. In this way, the human body quality and human body attributes of the human body in the image can be identified, and the human body attribute recognition result can be automatically corrected in combination with the result of human body quality recognition, without the cost of manual correction, thereby greatly saving the labor cost, improving the accuracy of human body attribute recognition, and improving the efficiency and effectiveness of human body attribute recognition, as well as the applicability of human body attribute recognition results.

[0101] 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.

[0102] In one embodiment, a human attribute detection device is provided, and the human attribute detection device corresponds one-to-one with the human attribute detection method in the above embodiment. Figure 6 As shown, the human attribute detection device includes an acquisition module 11, a recognition module 12, a correction module 13 and a determination module 14. The functional modules are described in detail as follows:

[0103] An acquisition module 11 is used to pre-process the image to be detected to obtain an image to be processed;

[0104] The recognition module 12 is used to perform human body quality and attribute recognition on the image to be processed through a human body attribute detection model to obtain a human body quality evaluation result and a human body attribute recognition result;

[0105] A correction module 13, configured to perform confidence correction processing on the human attribute recognition result according to the human quality evaluation result, to obtain a human recognition correction result;

[0106] The determination module 14 is used to generate a final result of human attributes corresponding to the image to be detected according to the human quality evaluation result and the human recognition correction result.

[0107] For the specific definition of the human attribute detection device, please refer to the definition of the human attribute detection method above, which will not be repeated here. Each module in the above human attribute detection 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.

[0108] 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 7 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, a human attribute detection method is implemented.

[0109] 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 human attribute detection method in the above embodiment is implemented.

[0110] 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 human attribute detection method in the above embodiment is implemented.

[0111] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed 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 can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an 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).

[0112] 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.

[0113] 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. A method for detecting human attributes, characterized in that: include: Preprocessing the image to be detected to obtain an image to be processed; Performing human body quality and attribute recognition on the image to be processed by using a human body attribute detection model to obtain a human body quality evaluation result and a human body attribute recognition result; According to the human body quality evaluation result, the human body attribute recognition result is processed for confidence correction to obtain a human body recognition correction result; Generate a final result of human attributes corresponding to the image to be detected according to the human body quality evaluation result and the human body recognition correction result; The method of performing human body quality and attribute recognition on the image to be processed by using a human body attribute detection model to obtain a human body quality evaluation result and a human body attribute recognition result includes: The shared features of the image to be processed are extracted through the backbone network to obtain a shared feature vector; the human attribute detection model includes the backbone network, a pooling layer, a human mass branch network and at least one human attribute branch network; Performing global average pooling processing on the shared feature vector through the pooling layer to obtain a shared feature map; The shared feature map is scored for human proportion by the human body quality branch network to obtain a proportion result, and the shared feature map is scored for human body occlusion to obtain an occlusion result, and the human body quality evaluation result is obtained according to the proportion result and the occlusion result; The shared feature graph is subjected to attribute recognition of respective branch tasks through each of the human attribute branch networks to obtain attribute branch results corresponding to each of the human attribute branch networks one by one, and all of the attribute branch results are merged into the human attribute recognition result.

2. The human attribute detection method according to claim 1, characterized in that: Before the image to be detected is preprocessed, the following steps are included: Acquire an input image, and input the input image into a single body recognition model; the input image contains at least one human body; Extracting monomer features from the input image using the monomer recognition model, performing monomer recognition based on the extracted monomer features, and identifying at least one monomer region containing a human body; An image corresponding to each of the monomer regions is extracted from the input image to obtain an image to be detected corresponding to each of the monomer regions one by one.

3. The human attribute detection method according to claim 1, characterized in that: The preprocessing of the image to be detected to obtain the image to be processed includes: Performing image enhancement processing on the image to be detected to obtain an enhanced image; Image channels of the image to be detected and the enhanced image are split and merged to obtain the image to be processed.

4. The human attribute detection method according to claim 3, characterized in that: The step of performing image enhancement processing on the image to be detected to obtain an enhanced image includes: Performing limb truncation recognition on the image to be detected by using a human body truncation detection model to obtain a truncation recognition result; When a truncation part is detected in the truncation recognition result, a truncation part prediction corresponding to the truncation part is performed on the image to be detected to obtain a prediction area, and image completion is performed on the image to be detected according to the prediction area, and image enhancement processing is performed on the image to be detected after image completion to obtain the enhanced image; When it is detected that there is no truncation part in the truncation recognition result, image enhancement processing is performed on the image to be detected to obtain the enhanced image.

5. The human attribute detection method according to claim 1, characterized in that: The step of scoring the human body proportion of the shared feature map by the human body mass branch network to obtain a proportion result includes: Upsampling the shared feature map through the human body mass branch network to obtain a map to be evaluated; The human body mass branch network is used to perform key limb point recognition, face recognition and body recognition on the image to be evaluated, and obtain key limb point coordinate results, face area results and body area results; Determine a first proportion value according to the distance between the key limb point coordinate result and the edge of the image to be detected, and determine a second proportion value according to the proportion of the face area result and the body area result in the image to be detected; The first proportion value and the second proportion value are weighted to obtain the proportion result.

6. The human attribute detection method according to claim 1, characterized in that: The human body quality evaluation result includes a quality score value and a quality confidence level; The step of performing confidence correction processing on the human attribute recognition result according to the human quality evaluation result to obtain a human recognition correction result includes: According to the quality confidence, each attribute branch result in the human attribute recognition result is subjected to confidence correction processing corresponding to the quality score value to obtain a branch correction result corresponding to each attribute branch result one by one; All the branch correction results are aggregated to obtain the human body recognition correction result.

7. A human attribute detection device, characterized in that: include: An acquisition module is used to preprocess the image to be detected to obtain an image to be processed; The recognition module is used to perform human body quality and attribute recognition on the image to be processed through a human body attribute detection model to obtain a human body quality evaluation result and a human body attribute recognition result; A correction module, used for performing confidence correction processing on the human attribute recognition result according to the human quality evaluation result to obtain a human recognition correction result; A determination module, used for generating a final result of human attributes corresponding to the image to be detected according to the human quality evaluation result and the human recognition correction result; The method of performing human body quality and attribute recognition on the image to be processed by using a human body attribute detection model to obtain a human body quality evaluation result and a human body attribute recognition result includes: The shared features of the image to be processed are extracted through the backbone network to obtain a shared feature vector; the human attribute detection model includes the backbone network, a pooling layer, a human mass branch network and at least one human attribute branch network; Performing global average pooling processing on the shared feature vector through the pooling layer to obtain a shared feature map; The shared feature map is scored for human proportion by the human body quality branch network to obtain a proportion result, and the shared feature map is scored for human body occlusion to obtain an occlusion result, and the human body quality evaluation result is obtained according to the proportion result and the occlusion result; The shared feature graph is subjected to attribute recognition of respective branch tasks through each of the human attribute branch networks to obtain attribute branch results corresponding to each of the human attribute branch networks one by one, and all of the attribute branch results are merged into the human attribute recognition result.

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 human attribute detection 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 human attribute detection method according to any one of claims 1 to 6 is implemented.

Images