WO2024011853A1 - 人体图像质量检测方法、装置、电子设备及存储介质 - Google Patents

人体图像质量检测方法、装置、电子设备及存储介质 Download PDF

Info

Publication number
WO2024011853A1
WO2024011853A1 PCT/CN2022/141457 CN2022141457W WO2024011853A1 WO 2024011853 A1 WO2024011853 A1 WO 2024011853A1 CN 2022141457 W CN2022141457 W CN 2022141457W WO 2024011853 A1 WO2024011853 A1 WO 2024011853A1
Authority
WO
WIPO (PCT)
Prior art keywords
human body
image
sample
body image
segmentation
Prior art date
Application number
PCT/CN2022/141457
Other languages
English (en)
French (fr)
Inventor
张洪
肖嵘
王孝宇
Original Assignee
青岛云天励飞科技有限公司
深圳云天励飞技术股份有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 青岛云天励飞科技有限公司, 深圳云天励飞技术股份有限公司 filed Critical 青岛云天励飞科技有限公司
Publication of WO2024011853A1 publication Critical patent/WO2024011853A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/74—Image or video pattern matching; Proximity measures in feature spaces
    • G06V10/75—Organisation of the matching processes, e.g. simultaneous or sequential comparisons of image or video features; Coarse-fine approaches, e.g. multi-scale approaches; using context analysis; Selection of dictionaries
    • G06V10/757—Matching configurations of points or features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks

Definitions

  • the present invention relates to the field of artificial intelligence, and in particular, to a human body image quality detection method, device, electronic equipment and storage medium.
  • the digital management of personnel mainly processes the visual information of personnel to form corresponding management solutions, such as archiving and managing captured images of personnel. Subsequent analysis and search of archived personnel images can improve the management efficiency of personnel information.
  • Archiving and managing captured images of people refers to archiving and archiving facial images and body images of captured individuals. Archive management is affected by image quality. The higher the image quality, the better the archive management effect.
  • existing methods generally use pedestrian re-identification and human body attribute recognition, and human body quality is a key factor affecting the effectiveness of pedestrian re-identification and human body attribute recognition.
  • Commonly used human body mass assessment methods are based on human body key points. However, in some images, the human body key points themselves may be inaccurate, which in turn leads to inaccurate human body mass assessment.
  • Embodiments of the present invention provide a human body image quality detection method, aiming to solve the existing problem of inaccurate human body quality assessment.
  • the human body quality detection model is trained to learn the implicit relationship between the human body image and the human body image quality, so that the human body quality detection model can output the quality detection of the target human body image based on the implicit relationship.
  • the training of the human body mass detection model is assisted by the sample human body image and the corresponding human body segmentation label, so that the human body mass detection model learns the correlation between human body segmentation and human body image quality, thereby improving the quality detection of the human body mass detection model. Results accuracy.
  • embodiments of the present invention provide a human body image quality detection method, which method includes:
  • an image to be detected where the image to be detected includes a target human body image
  • the human body quality detection model is trained based on a sample human body image, an image quality label corresponding to the sample human body image, and a human body segmentation label corresponding to the sample human body image.
  • the method before performing human body image quality detection on the image to be detected using a trained human body quality detection model to obtain a quality detection result of the target human body image, the method further includes:
  • the initial human body mass detection model is trained according to the training data set to obtain the trained human body mass detection model.
  • the sample image includes a plurality of sample human body images, the sample image is segmented to obtain a sample human body image, and the sample human body image is annotated according to the segmentation result to obtain the sample human body image.
  • Corresponding human body segmentation labels include:
  • first sample human body image is the human body image that is centered and has the largest area among the sample images
  • the first sample human body image and the second sample human body image are classified and labeled to obtain human body segmentation labels corresponding to the first sample human body image and the second sample human body image.
  • perform image quality labeling on the sample image to obtain a corresponding image quality label including:
  • an image quality label corresponding to the sample image is obtained.
  • the first sample human body image is image processed using a preset image processing method to obtain a third sample image, including:
  • the first sample human body image is truncated according to the preset truncation direction and truncation ratio to obtain a truncated human body image.
  • the truncation direction and the truncation ratio are also used to perform truncation image quality labeling on the truncated human body image. ;
  • the first sample human body image is subjected to occlusion processing according to the preset occlusion direction and occlusion ratio to obtain an occlusion human body image.
  • the occlusion direction and the occlusion ratio are also used to mark the occlusion image quality of the occluded human body image. ;
  • the first sample human body image is blurred according to the preset blur parameters to obtain a blurred human body image, and the blur parameters are also used to mark the blurred image quality of the blurred human body image;
  • a third sample image is obtained.
  • the initial human body mass detection model includes a backbone network, a human body segmentation branch network and a human body mass branch network.
  • the initial human body mass detection model is trained according to the training data set to obtain the training Good human body quality detection models include:
  • the parameters of the initial human body mass detection model are iteratively adjusted to obtain the trained human body mass detection model.
  • iteratively adjusting parameters of the initial human body mass detection model based on the first loss and the second loss to obtain the trained human body mass detection model includes:
  • the first loss and the second loss adjust the network parameters corresponding to the backbone network, the human body segmentation branch network and the human body mass branch network through stochastic gradient descent;
  • the training is stopped, and the trained human body segmentation branch network is deleted to obtain the trained human body quality detection Model.
  • a human body image quality detection device which includes:
  • the first acquisition module is used to acquire an image to be detected, where the image to be detected includes a target human body image;
  • the detection module is used to perform human body image quality detection on the image to be detected through the trained human body quality detection model, and obtain the quality detection result of the target human body image;
  • the human body quality detection model is trained based on a sample human body image, an image quality label corresponding to the sample human body image, and a human body segmentation label corresponding to the sample human body image.
  • embodiments of the present invention provide an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor.
  • the processor executes the computer program Implement the steps in the human body image quality detection method provided by the embodiment of the present invention.
  • embodiments of the present invention provide a computer-readable storage medium.
  • a computer program is stored on the computer-readable storage medium.
  • the human body image quality detection method provided by the embodiment of the invention is implemented. steps in.
  • an image to be detected is obtained, and the image to be detected includes a target human body image; the human body image quality detection is performed on the image to be detected through a trained human body quality detection model, and the quality detection result of the target human body image is obtained. ;
  • the human body quality detection model is trained based on a sample human body image, an image quality label corresponding to the sample human body image, and a human body segmentation label corresponding to the sample human body image. Through the sample human body image and the corresponding image quality label, the human body quality detection model is trained to learn the implicit relationship between the human body image and the human body image quality, so that the human body quality detection model can output the quality detection of the target human body image based on the implicit relationship.
  • the training of the human body mass detection model is assisted by the sample human body image and the corresponding human body segmentation label, so that the human body mass detection model learns the correlation between human body segmentation and human body image quality, thereby improving the quality detection of the human body mass detection model. Results accuracy.
  • Figure 1 is a flow chart of a human body image quality detection method provided by an embodiment of the present invention
  • Figure 2 is a schematic structural diagram of an initial human body mass detection model provided by an embodiment of the present invention.
  • Figure 3 is a schematic structural diagram of a human body image quality detection device provided by an embodiment of the present invention.
  • Figure 4 is a schematic structural diagram of an electronic device provided by an embodiment of the present invention.
  • Figure 1 is a flow chart of a human body image quality detection method provided by an embodiment of the present invention. As shown in Figure 1, the human body image quality detection method includes the following steps:
  • the image to be detected includes a target human body image, and the image to be detected may be a picture or a video.
  • the above-mentioned image to be detected may be uploaded by the user, or may be captured by an image collection device. It should be noted that the above-mentioned image to be detected may include one or more target human body images, and the above-mentioned target human body image may be understood as needing to be processed. Archived human body images, human body images can understand part of the image in the image to be detected, the image to be detected is a large image, and the target human body image is a small image in the large image.
  • the above target human body images can be understood as all human body images in the image to be detected, or specified human body images in the image to be detected.
  • the human body image quality can be used to evaluate the integrity and clarity of the human body in the image.
  • the more complete the human body in the image the higher the human body image quality.
  • the clearer the human body in the image the higher the human body image quality. high.
  • the image to be detected can be input into the trained human body quality detection model, the image to be detected is calculated and processed through the human body quality detection model, and the quality detection result of the corresponding target human body image is output.
  • the above human body quality detection model is trained based on the sample human body image, the image quality label corresponding to the sample human body image, and the human body segmentation label corresponding to the sample human body image.
  • the above human body mass detection model may be a human body mass detection model constructed based on a convolutional neural network.
  • the above image quality label is used to describe the real human body image quality of the sample human body image.
  • the image quality label can include labels of truncation, occlusion, blur level and other types.
  • the above-mentioned human body segmentation label is used to describe the real human body segmentation information of the sample human body image.
  • the above-mentioned human body segmentation label can be the position of the human body in the image area occupied by the sample human body image.
  • an image to be detected is obtained, and the image to be detected includes a target human body image; the human body image quality detection is performed on the image to be detected through a trained human body quality detection model, and the quality detection result of the target human body image is obtained. ;
  • the human body quality detection model is trained based on a sample human body image, an image quality label corresponding to the sample human body image, and a human body segmentation label corresponding to the sample human body image. Through the sample human body image and the corresponding image quality label, the human body quality detection model is trained to learn the implicit relationship between the human body image and the human body image quality, so that the human body quality detection model can output the quality detection of the target human body image based on the implicit relationship.
  • the training of the human body mass detection model is assisted by the sample human body image and the corresponding human body segmentation label, so that the human body mass detection model learns the correlation between human body segmentation and human body image quality, thereby improving the quality detection of the human body mass detection model. Results accuracy.
  • the above sample image may be obtained by uploading by the user, may be captured by an image collection device, or may be obtained by an image generation method.
  • the above-mentioned initial human body mass detection model may be a human body mass detection model constructed based on a convolutional neural network.
  • it can be a human body mass detection model based on convolutional neural networks such as ResNet, MobileNet, and ConvNeXt.
  • the sample image can be segmented through an image segmentation algorithm to obtain a sample human body image.
  • the above sample image can include one or more sample human body images.
  • the above sample human body image can be understood as a human body image that needs to be segmented.
  • the human body image can be understood as a part of the image in the sample image.
  • the sample image is a large image, and the sample human body image is a large image. Small picture in the picture.
  • the sample human body image can be annotated according to the location area of the sample human body image to obtain the human body segmentation label.
  • the corresponding label is a subject person label (the corresponding label value can be 1) or a non-subject person (the corresponding label value can be 0).
  • image quality annotation may include annotating the truncation ratio, occlusion ratio, blur degree and other types of the sample human body image to obtain labels of the truncation, occlusion, blur degree and other types of the sample image.
  • Each sample human body image corresponds to a sample image.
  • the human body segmentation label is associated with the sample image, and the image quality label is associated with the sample image to construct a training data set; the initial human body quality detection model is trained based on the training data set. Predict the initial human body mass detection model sample image to obtain the prediction result. Calculate the loss between the prediction result and the human body segmentation label and image quality label. Train and optimize the initial human body mass detection model by minimizing the loss to adjust the initial human body mass.
  • the model parameters of the detection model when the initial human body mass detection model converges or reaches the preset number of iterations, the trained human body mass detection model can be obtained.
  • Training the initial human body mass detection model through image quality labels can enable the initial human body mass detection model to learn the quality detection capabilities of human body images. By adding human body segmentation labels to sample images, it can assist the training of human body mass detection models and improve human body quality.
  • the detection model learns the correlation between human body segmentation and human body image quality. At the same time, it does not need to detect human body key points, thereby improving the accuracy of the quality detection results of the human body quality detection model.
  • the sample image includes multiple sample human body images, perform image segmentation on the sample image to obtain a sample human body image, and label the sample human body image according to the segmentation results to obtain a human body segmentation label corresponding to the sample human body image.
  • the sample image can also be image segmented to obtain the first sample human body image and the second sample human body image.
  • the first sample human body image is the human body image that is centered and has the largest area among the sample images; for the first sample human body image Classify and label the second sample human body image to obtain human body segmentation labels corresponding to the first sample human body image and the second sample human body image.
  • the sample image can be segmented through the existing human body image segmentation algorithm to segment the sample human body image in the sample image.
  • Each person's human body in the sample image corresponds to a sample human body image.
  • the above-mentioned first sample human body image may be a subject portrait in the sample image, and may also be called a subject person in the sample image.
  • the above-mentioned second sample human body image may be a non-subject portrait in the sample image, which may also be called a subject person in the sample image. non-subject person.
  • the above-mentioned first sample human body image may be a human body image that is centered and has the largest area in the sample image, and the human body image in the sample image is the second sample human body image. Label the first sample human body image and the second sample human body image to obtain human body segmentation labels.
  • the first sample human body image can be labeled as the subject person
  • the second sample human body image can be labeled as the non-subject person
  • the label corresponding to the first sample human body image is the subject person label (the corresponding label value can be 1)
  • the label corresponding to the second sample human body image is a non-subject person (the corresponding label value may be 0).
  • the first sample human body image can be image processed using a preset image processing method to obtain the third sample image; according to the preset image processing method, The image processing method is assumed to perform a first image quality label on the third sample image; calculate the area ratio between the second sample human body image and the first sample human body image; and perform a second image quality label on the sample image based on the area ratio. ; Based on the first image quality annotation and the second image quality annotation, obtain an image quality label corresponding to the sample image.
  • the first sample human body image can be extracted from the sample image, and the first sample human body image can be extracted through preset image processing
  • the method performs image processing on the first sample human body image, and after obtaining the third sample image, returns the third sample image to the sample image to form a new sample image.
  • the above image processing methods may include image truncation, image occlusion, image blur and other image processing methods.
  • the corresponding image processing method may be used as the third sample human body image.
  • the annotation information of the image is used to obtain the first image quality annotation. For example, if the upper body of the first sample human body image is truncated with an upper body truncation ratio of 20%, then the annotation information corresponding to the third sample image is an upper body truncation ratio of 20%, and the corresponding image quality label in the new sample image includes upper body truncation. Scale 20% label.
  • the above-mentioned area ratio can be the ratio of the sum of the areas of the second sample human body image and the area of the first sample human body image.
  • the first sample human body image and the second sample human body image can be obtained.
  • the area ratio is used as the second image quality label of the sample image, so that the image quality label corresponding to the sample image can be obtained based on the first image quality label and the second image quality label.
  • the first sample human body image can be processed according to the preset truncation direction and truncation ratio. Truncation processing is performed to obtain a truncated human body image.
  • the truncation direction and truncation ratio are also used to mark the truncated image quality of the truncated human body image; the first sample human body image is blocked according to the preset occlusion direction and occlusion ratio to obtain a blocked human body image.
  • the occlusion direction and occlusion ratio are also used to mark the occlusion image quality of the occluded human body image;
  • the first sample human body image is blurred according to the preset blur parameters to obtain a blurred human body image, and the blur parameters are also used to mark the blurred human body image.
  • Carry out blurred image quality annotation obtain the third sample image based on truncated human body images, occluded human body images and blurred human body images.
  • the above-mentioned third sample image may include truncated human body images, blocked human body images, and blurred human body images.
  • the above-mentioned image processing methods may include image processing methods such as image truncation, image occlusion, and image blurring.
  • image truncation includes a truncation direction and a truncation ratio.
  • the truncation direction can be up, down, left, and other directions.
  • the above truncation direction and truncation ratio can be randomly selected. By randomly selecting the truncation direction and the truncation ratio, the first sample human body image is truncated. Process to obtain the truncated human body image.
  • Image occlusion includes occlusion direction and occlusion ratio.
  • the occlusion direction can be up, down, left, right, etc.
  • the above occlusion direction and occlusion ratio can be randomly selected.
  • the above image blur may include blur type and blur degree.
  • the blur type may be motion blur, Gaussian blur, etc.
  • the first sample human body image is blurred to obtain a blurred human body image.
  • the data volume of the sample image can be increased, thereby improving the accuracy of the human body quality detection model.
  • the initial human body mass detection model includes a backbone network, a human body segmentation branch network and a human body mass branch network.
  • the common features of the sample images can be extracted through the backbone network; the common features are predicted through the human body segmentation branch network to obtain the human body segmentation prediction, and the common features are predicted through the human body mass branch network to obtain the image quality prediction; human body segmentation is calculated The first loss between the prediction and the human body segmentation label, and the second loss between the image quality prediction and the image quality label are calculated; based on the first loss and the second loss, iteratively adjust the parameters of the initial human body quality detection model to get Trained human body mass detection model.
  • the above-mentioned backbone network can be a convolutional neural network such as ResNet, MobileNet, ConvNeXt, etc.
  • the above-mentioned backbone network is used to extract common features of the human body segmentation branch network and the human body mass branch network.
  • the above-mentioned human body segmentation branch network includes an upsampling layer, and the above-mentioned human body segmentation branch network is used to extract implicit information corresponding to human body segmentation in the common features to obtain human body segmentation prediction of the sample image.
  • the above-mentioned human body mass branch network includes a pooling layer and a fully connected layer.
  • the above-mentioned pooling layer can be a global average pooling.
  • the number of the above-mentioned fully connected layers can be multiple.
  • the above-mentioned human body mass branch network is used to combine common features.
  • the implicit information corresponding to the human body image is extracted to obtain the image quality prediction of the sample image.
  • Figure 2 is a schematic structural diagram of an initial human body mass detection model provided by an embodiment of the present invention.
  • the output end of the backbone network is connected to the input end of the human body segmentation branch network.
  • the output end of the backbone network is connected to the input end of the human body mass branch network.
  • the sample image is processed through the backbone network to obtain a feature map of common features.
  • the common features are input to the human body segmentation branch network and the human body mass branch network respectively.
  • the human body segmentation branch network outputs the human segmentation prediction of the sample image
  • the human body quality branch network outputs the image quality prediction of the sample image.
  • the image quality prediction may include the truncation ratio, the occlusion ratio, the degree of blur, and the degree of multiple people.
  • the above-mentioned image quality prediction may include the upper body truncation ratio, the lower body truncation ratio, the left truncation ratio, the right truncation ratio, the upper body The proportion of being blocked, the proportion of the lower body being blocked, the proportion of being blocked on the left, the proportion of being blocked on the right, the degree of blur, and the degree of multiple people.
  • the above first loss is used to represent the degree of difference between human body segmentation prediction and human body segmentation label. If the first loss is smaller, it means that the smaller the difference between human body segmentation prediction and human body segmentation label, the more accurate the model prediction will be; if The larger the loss, the greater the difference between the human body segmentation prediction and the human body segmentation label, and the less accurate the model prediction.
  • the above-mentioned second loss is used to represent the degree of difference between the image quality prediction and the image quality label. If the second loss is smaller, it means that the smaller the difference between the image quality prediction and the image quality label, the more accurate the model prediction will be; if The larger the second loss is, the greater the difference between the image quality prediction and the image quality label, and the less accurate the model prediction is.
  • the initial human body mass detection model is iteratively trained with the first loss and the second loss as the minimum. Until the initial human body mass detection model converges or reaches the preset number of iterations, the training can be stopped and the trained human body can be obtained. Quality inspection model. By calculating the first loss between human segmentation prediction and human segmentation label, and combining the second loss between image quality prediction and image quality label to train the initial human body mass detection model, the training effect can be improved, thereby improving the human body mass detection model. accuracy.
  • the step of iteratively adjusting the parameters of the initial human body mass detection model based on the first loss and the second loss to obtain the trained human body mass detection model you can use randomization based on the first loss and the second loss.
  • Gradient descent adjusts the network parameters corresponding to the backbone network, human body segmentation branch network, and human body mass branch network; when the first loss and the second loss are minimum, or the number of iterations reaches the preset number, the training is stopped and the trained human body is Divide the branch network and delete it to obtain the trained human body quality detection model.
  • the above first loss can be calculated by a first loss function, and the above first loss function can be expressed as follows:
  • the above-mentioned Loss dice is the first loss
  • the above-mentioned predict is the human body segmentation prediction
  • the above-mentioned true is the human body segmentation label.
  • the above-mentioned second loss can be calculated by a second loss function.
  • the above-mentioned second loss function can be an L2 loss function.
  • the L2 loss function can also be called an L2 norm loss function, which is also called the least square error. In general, the sum of squares of the difference between the image quality label and the image quality prediction is minimized.
  • the first loss and the second loss can be added together to obtain the total loss.
  • the training is stopped and the trained human body segmentation branch network is deleted to obtain the well-trained Human body mass detection model.
  • the detection accuracy of the human body mass detection model can be improved.
  • the human body segmentation branch network can be deleted, reducing the amount of model data during deployment, and improving the deployment speed and running speed of the human body quality detection model.
  • human body image quality detection method provided by the embodiment of the present invention can be applied to smart phones, computers, servers and other devices that can perform human body image quality detection.
  • Figure 3 is a schematic structural diagram of a human body image quality detection device provided by an embodiment of the present invention. As shown in Figure 3, the device includes:
  • the first acquisition module 301 is used to acquire images to be detected, where the images to be detected include target human body images;
  • the detection module 302 is used to perform human body image quality detection on the image to be detected through the trained human body quality detection model, and obtain the quality detection result of the target human body image;
  • the human body quality detection model is trained based on a sample human body image, an image quality label corresponding to the sample human body image, and a human body segmentation label corresponding to the sample human body image.
  • the device also includes:
  • the second acquisition module is used to acquire sample images and initial human body quality detection models
  • the first annotation module is used to perform image segmentation on the sample image to obtain a sample human body image, and annotate the sample human body image according to the segmentation results to obtain a human body segmentation label corresponding to the sample human body image;
  • the second labeling module is used to label the sample image for image quality and obtain the corresponding image quality label
  • a construction module configured to construct a training data set based on the sample human body image, the human body segmentation label, and the image quality label;
  • a training module configured to train the initial human body mass detection model according to the training data set to obtain the trained human body mass detection model.
  • the sample images include multiple sample human body images
  • the first annotation module includes:
  • Segmentation submodule used to perform image segmentation on the sample image to obtain a first sample human body image and a second sample human body image.
  • the first sample human body image is the human body that is centered and has the largest area in the sample image. image;
  • the first labeling submodule is used to classify and label the first sample human body image and the second sample human body image, and obtain the human body segmentation corresponding to the first sample human body image and the second sample human body image. Label.
  • the second annotation module includes:
  • the first processing submodule is used to perform image processing on the first sample human body image through a preset image processing method to obtain a third sample image;
  • the second labeling submodule is used to perform first image quality labeling on the third sample image according to the preset image processing method
  • the first calculation sub-module is used to calculate the area ratio between the second sample human body image and the first sample human body image
  • a third labeling submodule configured to perform a second image quality label on the sample image according to the area ratio
  • the second processing submodule is used to obtain the image quality label corresponding to the sample image based on the first image quality label and the second image quality label.
  • the first processing sub-module includes:
  • a first processing unit configured to truncate the first sample human body image according to a preset truncation direction and truncation ratio to obtain a truncated human body image.
  • the truncation direction and the truncation ratio are also used to truncate the truncation image.
  • Human body images are truncated and image quality annotated;
  • the second processing unit is used to perform occlusion processing on the first sample human body image according to the preset occlusion direction and occlusion ratio to obtain an occlusion human body image.
  • the occlusion direction and the occlusion ratio are also used to perform occlusion processing on the occlusion human body image.
  • Human body images are marked for occlusion image quality;
  • the third processing unit is used to perform blur processing on the first sample human body image according to preset blur parameters to obtain a blurred human body image.
  • the blur parameters are also used to perform blur image quality annotation on the blurred human body image;
  • a fourth processing unit configured to obtain a third sample image based on the truncated human body image, the occluded human body image, and the blurred human body image.
  • the initial human body mass detection model includes a backbone network, a human body segmentation branch network and a human body mass branch network.
  • the training module includes:
  • An extraction submodule used to extract common features of the sample images through the backbone network
  • Prediction sub-module used to predict the common features through the human body segmentation branch network to obtain human body segmentation prediction, and predict the common features through the human body mass branch network to obtain image quality prediction;
  • a second calculation submodule configured to calculate the first loss between the human body segmentation prediction and the human body segmentation label, and calculate the second loss between the image quality prediction and the image quality label;
  • the adjustment sub-module is used to iteratively adjust parameters of the initial human body mass detection model according to the first loss and the second loss to obtain the trained human body mass detection model.
  • the adjustment sub-module includes:
  • An adjustment unit configured to adjust network parameters corresponding to the backbone network, the human body segmentation branch network and the human body mass branch network through stochastic gradient descent according to the first loss and the second loss;
  • a deletion unit configured to stop training when the first loss and the second loss are minimum, or when the number of iterations reaches a preset number, and delete the trained human body segmentation branch network to obtain the training Good human body mass detection model.
  • human body image quality detection device provided by the embodiment of the present invention can be applied to smart phones, computers, servers and other equipment that can perform human body image quality detection.
  • the human body image quality detection device provided by the embodiment of the present invention can realize each process implemented by the human body image quality detection method in the above method embodiment, and can achieve the same beneficial effects. To avoid repetition, they will not be repeated here.
  • Figure 4 is a schematic structural diagram of an electronic device provided by an embodiment of the present invention. As shown in Figure 4, it includes: a memory 402, a processor 401, and an electronic device stored in the memory 402 and available in the processor.
  • a computer program for a human body image quality detection method running on 401 wherein:
  • the processor 401 is used to call the computer program stored in the memory 402 and perform the following steps:
  • an image to be detected where the image to be detected includes a target human body image
  • the human body quality detection model is trained based on a sample human body image, an image quality label corresponding to the sample human body image, and a human body segmentation label corresponding to the sample human body image.
  • the method executed by the processor 401 also includes:
  • the initial human body mass detection model is trained according to the training data set to obtain the trained human body mass detection model.
  • the sample image includes a plurality of sample human body images.
  • the processor 401 performs image segmentation on the sample image to obtain a sample human body image, and labels the sample human body image according to the segmentation results to obtain
  • the human body segmentation labels corresponding to the sample human body images include:
  • first sample human body image is the human body image that is centered and has the largest area among the sample images
  • the first sample human body image and the second sample human body image are classified and labeled to obtain human body segmentation labels corresponding to the first sample human body image and the second sample human body image.
  • the processor 401 performs image quality labeling on the sample image to obtain the corresponding image quality label, including:
  • an image quality label corresponding to the sample image is obtained.
  • the processor 401 performs image processing on the first sample human body image through a preset image processing method to obtain a third sample image, including:
  • the first sample human body image is truncated according to the preset truncation direction and truncation ratio to obtain a truncated human body image.
  • the truncation direction and the truncation ratio are also used to perform truncation image quality labeling on the truncated human body image. ;
  • the first sample human body image is subjected to occlusion processing according to the preset occlusion direction and occlusion ratio to obtain an occlusion human body image.
  • the occlusion direction and the occlusion ratio are also used to mark the occlusion image quality of the occluded human body image. ;
  • a third sample image is obtained.
  • the initial human body mass detection model includes a backbone network, a human body segmentation branch network and a human body mass branch network.
  • the processor 401 executes the training of the initial human body mass detection model based on the training data set. , obtain the trained human body quality detection model, including:
  • the parameters of the initial human body mass detection model are iteratively adjusted to obtain the trained human body mass detection model.
  • the processor 401 performs iterative adjustment of parameters of the initial human body mass detection model based on the first loss and the second loss to obtain the trained human body mass detection model, including :
  • the first loss and the second loss adjust the network parameters corresponding to the backbone network, the human body segmentation branch network and the human body mass branch network through stochastic gradient descent;
  • the training is stopped, and the trained human body segmentation branch network is deleted to obtain the trained human body quality detection Model.
  • the electronic device provided by the embodiment of the present invention can implement each process implemented by the human body image quality detection method in the above method embodiment, and can achieve the same beneficial effects. To avoid repetition, they will not be repeated here.
  • Embodiments of the present invention also provide a computer-readable storage medium.
  • a computer program is stored on the computer-readable storage medium.
  • the computer program is executed by a processor, the human body image quality detection method or the application-end human body image provided by the embodiment of the present invention is implemented.
  • Each process of the quality inspection method can achieve the same technical effect. To avoid duplication, it will not be described again here.
  • the program can be stored in a computer-readable storage medium.
  • the program can be stored in a computer-readable storage medium.
  • the process may include the processes of the embodiments of each of the above methods.
  • the storage medium can be a magnetic disk, an optical disk, a read-only memory (Read-Only Memory, ROM) or a random access memory (Random Access Memory, referred to as RAM), etc.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Multimedia (AREA)
  • General Physics & Mathematics (AREA)
  • Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Computation (AREA)
  • Databases & Information Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Medical Informatics (AREA)
  • Software Systems (AREA)
  • Computing Systems (AREA)
  • Artificial Intelligence (AREA)
  • Health & Medical Sciences (AREA)
  • Human Computer Interaction (AREA)
  • Image Analysis (AREA)

Abstract

本发明实施例提供一种人体图像质量检测方法，方法包括：获取待检测图像，待检测图像包括目标人体图像；通过训练好的人体质量检测模型对的待检测图像进行人体图像质量检测，得到目标人体图像的质量检测结果；其中，人体质量检测模型根据样本人体图像、与样本人体图像对应的图像质量标签以及与样本人体图像对应的人体分割标签进行训练得到。通过训练人体质量检测模型学习到人体图像与人体图像质量之间的隐性关系，同时，通过样本人体图像和对应的人体分割标签辅助人体质量检测模型的训练，使得人体质量检测模型学习到人体分割与人体图像质量之间的关联关系，从而提高人体质量检测模型的质量检测结果准确性。

Description

人体图像质量检测方法、装置、电子设备及存储介质
本申请要求于2022年7月12日提交中国专利局，申请号为202210819601.X、发明名称为“人体图像质量检测方法、装置、电子设备及存储介质”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本发明涉及本发明涉及人工智能领域，尤其涉及一种人体图像质量检测方法、装置、电子设备及存储介质。
背景技术
人员作为数字城市中重要的组成单元，对人员进行数字化管理是数字城市中必不可少的部分。人员的数字化管理主要通过对人员的视觉信息进行处理，形成对应的管理方案，比如对人员抓拍图像进行归档管理，通过归档的人员图像进行后续分析和查找，可以提高人员信息的管理效率。对人员抓拍图像进行归档管理指的是抓拍人员的人脸图像和人体图像进行建档和归档，归档管理受图像质量的影响，图像质量越高，则归档管理的效果越好。对于人体图像的归档，现有方法一般采用行人重识别和人体属性识别，而人体质量是影响行人重识别、人体属性识别效果的关键因素。常用的人体质量评估方法是基于人体关键点进行评估的，然而，在一些图像中人体关键点本身就可能是不准确的，进而导致人体质量评估不准确。
发明内容
本发明实施例提供一种人体图像质量检测方法，旨在解决现有人体质量评估不准确的问题。通过样本人体图像和对应的图像质量标签，训练人体质量检测模型学习到人体图像与人体图像质量之间的隐性关系，从而使得人体质量检测模型能够根据该隐性关系输出目标人体图像的质量检测结果，同时，通过样本人体图像和对应的人体分割标签辅助人体质量检测模型的训练，使得人体质量检测模型学习到人体分割与人体图像质量之间的关联关系，从而提高人体质量检测模型的质量检测结果准确性。
第一方面，本发明实施例提供一种人体图像质量检测方法，所述方法包括：
获取待检测图像，所述待检测图像包括目标人体图像；
通过训练好的人体质量检测模型对的待检测图像进行人体图像质量检测，得到所述目标人体图像的质量检测结果；
其中，所述人体质量检测模型根据样本人体图像、与所述样本人体图像对应的图像质量标签以及与所述样本人体图像对应的人体分割标签进行训练得 到。
可选的，在所述通过训练好的人体质量检测模型对所述待检测图像进行人体图像质量检测，得到所述目标人体图像的质量检测结果之前，所述方法还包括：
获取样本图像以及初始人体质量检测模型；
对所述样本图像进行图像分割，得到样本人体图像，并根据分割结果对所述样本人体图像进行标注，得到所述样本人体图像对应的人体分割标签；
对所述样本图像进行图像质量标注，得到对应的图像质量标签；
基于所述样本人体图像、所述人体分割标签以及所述图像质量标签，构建得到训练数据集；
根据所述训练数据集对所述初始人体质量检测模型进行训练，得到所述训练好的人体质量检测模型。
可选的，所述样本图像包括多个样本人体图像，所述对所述样本图像进行图像分割，得到样本人体图像，并根据分割结果对所述样本人体图像进行标注，得到所述样本人体图像对应的人体分割标签，包括：
对所述样本图像进行图像分割，得到第一样本人体图像和第二样本人体图像，所述第一样本人体图像为在所述样本图像中居中且面积最大的人体图像；
对所述第一样本人体图像和所述第二样本人体图像进行分类标注，得到所述第一样本人体图像和所述第二样本人体图像对应的人体分割标签。
可选的，所述对所述样本图像进行图像质量标注，得到对应的图像质量标签，包括：
通过预设的图像处理方法对所述第一样本人体图像进行图像处理，得到第三样本图像；
根据所述预设的图像处理方法，对所述第三样本图像进行第一图像质量标注；
计算所述第二样本人体图像与所述第一样本人体图像之间的面积比；
根据所述面积比，对所述样本图像进行第二图像质量标注；
基于所述第一图像质量标注与所述第二图像质量标注，得到所述样本图像对应的图像质量标签。
可选的，所述通过预设的图像处理方法对所述第一样本人体图像进行图像处理，得到第三样本图像，包括：
根据预设的截断方向和截断比例对所述第一样本人体图像进行截断处理，得到截断人体图像，所述截断方向和所述截断比例还用于对所述截断人体图像进行截断图像质量标注；
根据预设的遮挡方向和遮挡比例对所述第一样本人体图像进行遮挡处理，得到遮挡人体图像，所述遮挡方向和所述遮挡比例还用于对所述遮挡人体图像进行遮挡图像质量标注；
根据预设的模糊参数对所述第一样本人体图像进行模糊处理，得到模糊人 体图像，所述模糊参数还用于对所述模糊人体图像进行模糊图像质量标注；
基于所述截断人体图像、所述遮挡人体图像以及所述模糊人体图像，得到第三样本图像。
可选的，所述初始人体质量检测模型包括主干网络、人体分割支路网络以及人体质量支路网络，所述根据所述训练数据集对所述初始人体质量检测模型进行训练，得到所述训练好的人体质量检测模型，包括：
通过所述主干网络提取所述样本图像的共用特征；
通过所述人体分割支路网络对所述共用特征进行预测，得到人体分割预测，以及通过所述人体质量支路网络对所述共用特征进行预测，得到图像质量预测；
计算所述人体分割预测与所述人体分割标签之间的第一损失，以及计算所述图像质量预测与所述图像质量标签之间的第二损失；
根据所述第一损失与所述第二损失，对所述初始人体质量检测模型进行参数的迭代调整，得到所述训练好的人体质量检测模型。
可选的，所述根据所述第一损失与所述第二损失，对所述初始人体质量检测模型进行参数的迭代调整，得到所述训练好的人体质量检测模型，包括：
根据所述第一损失与所述第二损失，通过随机梯度下降调整所述主干网络、所述人体分割支路网络以及所述人体质量支路网络对应的网络参数；
当所述第一损失与所述第二损失最小，或者迭代次数达到预设次数时，停止训练，并将训练好的所述人体分割支路网络进行删除，得到所述训练好的人体质量检测模型。
第二方面，本发明实施例提供一种人体图像质量检测装置，所述装置包括：
第一获取模块，用于获取待检测图像，所述待检测图像包括目标人体图像；
检测模块，用于通过训练好的人体质量检测模型对的待检测图像进行人体图像质量检测，得到所述目标人体图像的质量检测结果；
其中，所述人体质量检测模型根据样本人体图像、与所述样本人体图像对应的图像质量标签以及与所述样本人体图像对应的人体分割标签进行训练得到。
第三方面，本发明实施例提供一种电子设备，包括：存储器、处理器及存储在所述存储器上并可在所述处理器上运行的计算机程序，所述处理器执行所述计算机程序时实现本发明实施例提供的人体图像质量检测方法中的步骤。
第四方面，本发明实施例提供一种计算机可读存储介质，所述计算机可读存储介质上存储有计算机程序，所述计算机程序被处理器执行时实现发明实施例提供的人体图像质量检测方法中的步骤。
本发明实施例中，获取待检测图像，所述待检测图像包括目标人体图像；通过训练好的人体质量检测模型对的待检测图像进行人体图像质量检测，得到所述目标人体图像的质量检测结果；其中，所述人体质量检测模型根据样本人体图像、与所述样本人体图像对应的图像质量标签以及与所述样本人体图像对应的人体分割标签进行训练得到。通过样本人体图像和对应的图像质量标签， 训练人体质量检测模型学习到人体图像与人体图像质量之间的隐性关系，从而使得人体质量检测模型能够根据该隐性关系输出目标人体图像的质量检测结果，同时，通过样本人体图像和对应的人体分割标签辅助人体质量检测模型的训练，使得人体质量检测模型学习到人体分割与人体图像质量之间的关联关系，从而提高人体质量检测模型的质量检测结果准确性。
附图说明
下面将对本申请实施例中所需要使用的附图作介绍。
图1是本发明实施例提供的一种人体图像质量检测方法的流程图；
图2是本发明实施例提供的一种初始人体质量检测模型的结构示意图；
图3是本发明实施例提供的一种人体图像质量检测装置的结构示意图；
图4是本发明实施例提供的一种电子设备的结构示意图。
具体实施方式
下面结合附图对本申请的实施例进行描述。
请参见图1，图1是本发明实施例提供的一种人体图像质量检测方法的流程图，如图1所示，该人体图像质量检测方法包括以下步骤：
101、获取待检测图像。
在本发明实施例中，上述待检测图像包括目标人体图像，上述待检测图像可以是图片或视频。上述待检测图像可以是通过用户上传得到，也可以是通过图像采集设备抓拍得到，需要说明的是，上述待检测图像中可以包括一个或多个目标人体图像，上述目标人体图像可以理解为需要进行归档的人体图像，人体图像可以理解待检测图像中的一部分图像，待检测图像为大图，目标人体图像为大图中的小图。
上述目标人体图像可以理解为待检测图像中的所有人体图像，也可以待检测图像中指定的人体图像。
102、通过训练好的人体质量检测模型对的待检测图像进行人体图像质量检测，得到目标人体图像的质量检测结果。
在本发明实施例中，人体图像质量可以用于评估图像中人体完整度和清晰度，图像中人体越完整，则人体图像质量越高，同样的，图像中人体越清晰，则人体图像质量越高。
具体的，可以将待检测图像输入到训练好的人体质量检测模型中，通过人体质量检测模型对待检测图像进行计算处理，输出对应的目标人体图像的质量检测结果。
进一步的，上述人体质量检测模型根据样本人体图像、与样本人体图像对应的图像质量标签以及与样本人体图像对应的人体分割标签进行训练得到。上述人体质量检测模型可以是基于卷积神经网络进行构建的人体质量检测模型。
上述图像质量标签用于描述样本人体图像的真实人体图像质量，具体来说， 图像质量标签可以包括截断、遮挡、模糊程度等类型的标签。上述人体分割标签用于描述样本人体图像的真实人体分割信息，上述人体分割标签可以是人体在样本人体图像中所占图像区域的位置。
本发明实施例中，获取待检测图像，所述待检测图像包括目标人体图像；通过训练好的人体质量检测模型对的待检测图像进行人体图像质量检测，得到所述目标人体图像的质量检测结果；其中，所述人体质量检测模型根据样本人体图像、与所述样本人体图像对应的图像质量标签以及与所述样本人体图像对应的人体分割标签进行训练得到。通过样本人体图像和对应的图像质量标签，训练人体质量检测模型学习到人体图像与人体图像质量之间的隐性关系，从而使得人体质量检测模型能够根据该隐性关系输出目标人体图像的质量检测结果，同时，通过样本人体图像和对应的人体分割标签辅助人体质量检测模型的训练，使得人体质量检测模型学习到人体分割与人体图像质量之间的关联关系，从而提高人体质量检测模型的质量检测结果准确性。
可选的，在通过训练好的人体质量检测模型对待检测图像进行人体图像质量检测，得到目标人体图像的质量检测结果之前，还可以获取样本图像以及初始人体质量检测模型；对样本图像进行图像分割，得到样本人体图像，并根据分割结果对样本人体图像进行标注，得到样本人体图像对应的人体分割标签；对样本图像进行图像质量标注，得到对应的图像质量标签；基于样本人体图像、人体分割标签以及图像质量标签，构建得到训练数据集；根据训练数据集对初始人体质量检测模型进行训练，得到训练好的人体质量检测模型。
在本发明实施例中，上述样本图像可以是通过用户上传得到，也可以是通过图像采集设备抓拍得到，还可以是通过图像生成方法得到。
上述初始人体质量检测模型可以是基于卷积神经网络进行构建的人体质量检测模型。比如，可以是基于ResNet、MobileNet、ConvNeXt等卷积神经网络构建得到的人体质量检测模型。
进一步的，可以通过图像分割算法对样本图像进行分割，得到样本人体图像。上述样本图像中可以包括一个或多个样本人体图像，上述样本人体图像可以理解为需要进行分割的人体图像，人体图像可以理解样本图像中的一部分图像，样本图像为大图，样本人体图像为大图中的小图。在得到样本人体图像后，可以根据样本人体图像的位置区域，对样本人体图像进行标注，得到人体分割标签。比如，将样本人体图像标注为主体人或非主体人，则对应的标签为主体人标签(对应标签值可以是1)或非主体人(对应标签值可以是0)。
进一步的，可以采用现有的质量评估算法或采用通过专家人工对样本图像进行图像质量标注，得到对应的图像质量标签。上述图像质量标注可以包括对样本人体图像的截断比例、遮挡比例、模糊程度等类型进行标注，得到样本图像的截断、遮挡、模糊程度等类型的标签。
每个样本人体图像对应一个样本图像，将人体分割标签与样本图像进行关联，以及将图像质量标签与样本图像进行关联，构建得到训练数据集；根据训 练数据集对初始人体质量检测模型进行训练，通过初始人体质量检测模型样本图像进行预测，得到预测结果，计算预测结果与人体分割标签以及图像质量标签之间的损失，通过最小化该损失对初始人体质量检测模型进行训练优化，调整初始人体质量检测模型的模型参数，当初始人体质量检测模型收敛或达到预设的迭代次数时，可以得到训练好的人体质量检测模型。
通过图像质量标签对初始人体质量检测模型进行训练，可以使得初始人体质量检测模型学习到人体图像的质量检测能力，通过对样本图像增加人体分割标签，可以辅助人体质量检测模型的训练，使得人体质量检测模型学习到人体分割与人体图像质量之间的关联关系，同时，不需要进行人体关键点检测，从而提高人体质量检测模型的质量检测结果准确性。
可选的，样本图像包括多个样本人体图像，对所述样本图像进行图像分割，得到样本人体图像，并根据分割结果对样本人体图像进行标注，得到样本人体图像对应的人体分割标签的步骤中，还可以对样本图像进行图像分割，得到第一样本人体图像和第二样本人体图像，第一样本人体图像为在样本图像中居中且面积最大的人体图像；对第一样本人体图像和第二样本人体图像进行分类标注，得到第一样本人体图像和第二样本人体图像对应的人体分割标签。
在本发明实施例中，可以通过现有的人体图像分割算法对样本图像进行图像分割，将样本图像中的样本人体图像分割出来，样本图像中每个人员的人体对应一个样本人体图像。
上述第一样本人体图像可以是样本图像中的主体人像，也可以称为样本图像中的主体人，上述第二样本人体图像可以是样本图像中的非主体人像，也可以称为样本图像中的非主体人。上述第一样本人体图像可以是在样本图像中居中且面积最大的人体图像，其中在样本图像中的人体图像则为第二样本人体图像。对第一样本人体图像和第二样本人体图像进行标注，得到人体分割标签。比如，可以将第一样本人体图像标注为主体人，将第二样本人体图像标注为非主体人，则第一样本人体图像对应的标签为主体人标签(对应标签值可以是1)，第二样本人体图像对应的标签为非主体人(对应标签值可以是0)。
通过对样本图像增加人体分割标签，可以辅助人体质量检测模型的训练，使得人体质量检测模型学习到人体分割与人体图像质量之间的关联关系，同时，不需要进行人体关键点检测，从而提高人体质量检测模型的质量检测结果准确性。
可选的，在对样本图像进行图像质量标注，得到对应的图像质量标签的步骤中，可以通过预设的图像处理方法对第一样本人体图像进行图像处理，得到第三样本图像；根据预设的图像处理方法，对第三样本图像进行第一图像质量标注；计算第二样本人体图像与第一样本人体图像之间的面积比；根据面积比，对样本图像进行第二图像质量标注；基于第一图像质量标注与第二图像质量标注，得到样本图像对应的图像质量标签。
在本发明实施例中，通过图像分割算法在样本图像中分割出第一样本人体 图像和第二样本人体图像后，可以在样本图像中提取第一样本人体图像，通过预设的图像处理方法对第一样本人体图像进行图像处理，得到第三样本图像后，将第三样本图像返回到样本图像中，形成新的样本图像。
上述图像处理方法可以包括图像截断、图像遮挡以及图像模糊等图像处理方法，在对第一样本人体图像进行图像处理，得到第三样本图像后，可以将对应的图像处理方法作为第三样本人体图像的标注信息，得到第一图像质量标注。比如以上身截断比例为20％对第一样本人体图像的上身进行截断，则对应第三样本图像的标注信息为上身截断比例20％，新的样本图像中对应的图像质量标签中包括上身截断比例20％的标签。
上述面积比可以是第二样本人体图像的面积之和与第一样本人体图像的面积的比值，通过面积比表示样本图像的多人程度，可以得到第一样本人体图像与第二样本人体图像的分布关系，面积比越大，则说明第二样本人体图像所占的图像面积越大，第一样本人体图像所占的图像面积越小，图像质量越差；面积比越小，则说明第二样本人体图像所占的图像面积越小，第一样本人体图像所占的图像面积越大，图像质量越好。将面积比作为样本图像的第二图像质量标注，从而可以根据第一图像质量标注和第二图像质量标注，得到样本图像对应的图像质量标签。
可选的，在通过预设的图像处理方法对第一样本人体图像进行图像处理，得到第三样本图像的步骤中，可以根据预设的截断方向和截断比例对第一样本人体图像进行截断处理，得到截断人体图像，截断方向和截断比例还用于对截断人体图像进行截断图像质量标注；根据预设的遮挡方向和遮挡比例对第一样本人体图像进行遮挡处理，得到遮挡人体图像，遮挡方向和遮挡比例还用于对遮挡人体图像进行遮挡图像质量标注；根据预设的模糊参数对第一样本人体图像进行模糊处理，得到模糊人体图像，模糊参数还用于对模糊人体图像进行模糊图像质量标注；基于截断人体图像、遮挡人体图像以及模糊人体图像，得到第三样本图像。
在本发明实施例中，上述第三样本图像可以包括截断人体图像、遮挡人体图像以及模糊人体图像，上述图像处理方法可以包括图像截断、图像遮挡以及图像模糊等图像处理方法。具体的，图像截断包括截断方向和截断比例，截断方向可以是上下左右等方向，上述截断方向和截断比例可以是随机选择，通过随机选择截断方向和截断比例，对第一样本人体图像进行截断处理，得到截断人体图像。图像遮挡包括遮挡方向和遮挡比例，遮挡方向可以是上下左右等方向，上述遮挡方向和遮挡比例可以是随机选择，通过随机选择遮挡方向和遮挡比例，对第一样本人体图像进行遮挡处理，得到遮挡人体图像。上述图像模糊可以包括模糊类型和模糊程度，上述模糊类型可以是运动模糊、高斯模糊等，通过随机选择模糊类型和模糊程度，对第一样本人体图像进行模糊处理，得到模糊人体图像。
通过对第一样本人体图像进行图像处理，可以增加样本图像的数据量，从 而提高人体质量检测模型的准确性。
可选的，初始人体质量检测模型包括主干网络、人体分割支路网络以及人体质量支路网络，在根据训练数据集对初始人体质量检测模型进行训练，得到训练好的人体质量检测模型的步骤中，可以通过主干网络提取样本图像的共用特征；通过人体分割支路网络对共用特征进行预测，得到人体分割预测，以及通过人体质量支路网络对共用特征进行预测，得到图像质量预测；计算人体分割预测与人体分割标签之间的第一损失，以及计算图像质量预测与图像质量标签之间的第二损失；根据第一损失与第二损失，对初始人体质量检测模型进行参数的迭代调整，得到训练好的人体质量检测模型。
在本发明实施例中，上述主干网络可以是ResNet、MobileNet、ConvNeXt等卷积神经网络，上述主干网络用于提取人体分割支路网络以及人体质量支路网络的共用特征。上述人体分割支路网络中包括上采样层，上述人体分割支路网络用于将共用特征中与人体分割相对应的隐含信息进行提取，得到样本图像的人体分割预测。上述人体质量支路网络中包括池化层和全连接层，上述池化层可以是全局平均池化，上述全连接层的数量可以是多个，通过上述人体质量支路网络用于将共用特征中与人体图像对应的隐含信息进行提取，得到样本图像的图像质量预测。
具体的，请参见图2，图2是本发明实施例提供的一种初始人体质量检测模型的结构示意图，如图2所示，主干网络的输出端与人体分割支路网络的输入端连接，同时主干网络的输出端与人体质量支路网络的输入端连接，样本图像通过主干网络处理，得到共用特征的特征图，将该共用特征分别输入到人体分割支路网络和人体质量支路网络，其中，人体分割支路网络输出样本图像的人体分割预测，人体质量支路网络输出样本图像的图像质量预测。其中，图像质量预测可以包括截断比例、遮挡比例、模糊程度、多人程度，具体的，上述图像质量预测可以包括上身被截断比例、下身被截断比例、左边被截断比例、右边被截断比例、上身被遮挡比例、下身被遮挡比例、左边被遮挡比例、右边被遮挡比例、模糊程度以及多人程度。
上述第一损失用于表示人体分割预测与人体分割标签之间相差程度，若第一损失越小，则说明人体分割预测与人体分割标签之间的相差程度越小，模型预测越准确；若第一损失越大，则说明人体分割预测与人体分割标签之间的相差程度越大，模型预测越不准确。上述第二损失用于表示图像质量预测与图像质量标签之间的相差程度，若第二损失越小，则说明图像质量预测与图像质量标签之间的相差程度越小，模型预测越准确；若第二损失越大，则说明图像质量预测与图像质量标签之间的相差程度越大，模型预测越不准确。
在训练过程中，以第一损失和第二损失为最小对初始人体质量检测模型进行迭代训练，直到初始人体质量检测模型收敛或达到预设的迭代次数，则可以停止训练，得到训练好的人体质量检测模型。通过计算人体分割预测与人体分割标签之间的第一损失，结合图像质量预测与图像质量标签之间的第二损失对 初始人体质量检测模型进行训练，可以提高训练效果，从而提高人体质量检测模型的准确性。
可选的，在根据第一损失与第二损失，对初始人体质量检测模型进行参数的迭代调整，得到训练好的人体质量检测模型的步骤中，可以根据第一损失与第二损失，通过随机梯度下降调整主干网络、人体分割支路网络以及人体质量支路网络对应的网络参数；当第一损失与第二损失最小，或者迭代次数达到预设次数时，停止训练，并将训练好的人体分割支路网络进行删除，得到训练好的人体质量检测模型。
在本发明实施例中，上述第一损失可以通过第一损失函数计算得到，上述第一损失函数可以如下述式子所示：
Figure PCTCN2022141457-appb-000001
其中，上述Loss dice为第一损失，上述predict为人体分割预测，上述true为人体分割标签。
上述第二损失可以通过第二损失函数计算得到，上述第二损失函数可以是L2损失函数，L2损失函数也可以称为L2范数损失函数，也被称为最小平方误差。总的来说，是把与图像质量标签与图像质量预测的差值的平方和最小化。
可以将第一损失和第二损失进行相加，得到总损失，当总损失最小，或者迭代次数达到预设次数时，停止训练，并将训练好的人体分割支路网络进行删除，得到训练好的人体质量检测模型。
通过人体分割支路网络辅助人体质量检测模型的训练，可以提高人体质量检测模型的检测准确性，在进行部署时，由于只需要输出对应的质量检测结果，而不需要输出对应的人体分割预测结果，因此，可以将人体分割支路网络进行删除，降低部署时的模型数据量，提高人体质量检测模型的部署速度和运行速度。
需要说明的是，本发明实施例提供的人体图像质量检测方法可以应用于可以进行人体图像质量检测的智能手机、电脑、服务器等设备。
可选的，请参见图3，图3是本发明实施例提供的一种人体图像质量检测装置的结构示意图，如图3所示，所述装置包括：
第一获取模块301，用于获取待检测图像，所述待检测图像包括目标人体图像；
检测模块302，用于通过训练好的人体质量检测模型对的待检测图像进行人体图像质量检测，得到所述目标人体图像的质量检测结果；
其中，所述人体质量检测模型根据样本人体图像、与所述样本人体图像对应的图像质量标签以及与所述样本人体图像对应的人体分割标签进行训练得到。
可选的，所述装置还包括：
第二获取模块，用于获取样本图像以及初始人体质量检测模型；
第一标注模块，用于对所述样本图像进行图像分割，得到样本人体图像，并根据分割结果对所述样本人体图像进行标注，得到所述样本人体图像对应的人体分割标签；
第二标注模块，用于对所述样本图像进行图像质量标注，得到对应的图像质量标签；
构建模块，用于基于所述样本人体图像、所述人体分割标签以及所述图像质量标签，构建得到训练数据集；
训练模块，用于根据所述训练数据集对所述初始人体质量检测模型进行训练，得到所述训练好的人体质量检测模型。
可选的，所述样本图像包括多个样本人体图像，所述第一标注模块包括：
分割子模块，用于对所述样本图像进行图像分割，得到第一样本人体图像和第二样本人体图像，所述第一样本人体图像为在所述样本图像中居中且面积最大的人体图像；
第一标注子模块，用于对所述第一样本人体图像和所述第二样本人体图像进行分类标注，得到所述第一样本人体图像和所述第二样本人体图像对应的人体分割标签。
可选的，所述第二标注模块包括：
第一处理子模块，用于通过预设的图像处理方法对所述第一样本人体图像进行图像处理，得到第三样本图像；
第二标注子模块，用于根据所述预设的图像处理方法，对所述第三样本图像进行第一图像质量标注；
第一计算子模块，用于计算所述第二样本人体图像与所述第一样本人体图像之间的面积比；
第三标注子模块，用于根据所述面积比，对所述样本图像进行第二图像质量标注；
第二处理子模块，用于基于所述第一图像质量标注与所述第二图像质量标注，得到所述样本图像对应的图像质量标签。
可选的，所述第一处理子模块包括：
第一处理单元，用于根据预设的截断方向和截断比例对所述第一样本人体图像进行截断处理，得到截断人体图像，所述截断方向和所述截断比例还用于对所述截断人体图像进行截断图像质量标注；
第二处理单元，用于根据预设的遮挡方向和遮挡比例对所述第一样本人体图像进行遮挡处理，得到遮挡人体图像，所述遮挡方向和所述遮挡比例还用于对所述遮挡人体图像进行遮挡图像质量标注；
第三处理单元，用于根据预设的模糊参数对所述第一样本人体图像进行模糊处理，得到模糊人体图像，所述模糊参数还用于对所述模糊人体图像进行模 糊图像质量标注；
第四处理单元，用于基于所述截断人体图像、所述遮挡人体图像以及所述模糊人体图像，得到第三样本图像。
可选的，所述初始人体质量检测模型包括主干网络、人体分割支路网络以及人体质量支路网络，所述训练模块，包括：
提取子模块，用于通过所述主干网络提取所述样本图像的共用特征；
预测子模块，用于通过所述人体分割支路网络对所述共用特征进行预测，得到人体分割预测，以及通过所述人体质量支路网络对所述共用特征进行预测，得到图像质量预测；
第二计算子模块，用于计算所述人体分割预测与所述人体分割标签之间的第一损失，以及计算所述图像质量预测与所述图像质量标签之间的第二损失；
调整子模块，用于根据所述第一损失与所述第二损失，对所述初始人体质量检测模型进行参数的迭代调整，得到所述训练好的人体质量检测模型。
可选的，所述调整子模块包括：
调整单元，用于根据所述第一损失与所述第二损失，通过随机梯度下降调整所述主干网络、所述人体分割支路网络以及所述人体质量支路网络对应的网络参数；
删除单元，用于当所述第一损失与所述第二损失最小，或者迭代次数达到预设次数时，停止训练，并将训练好的所述人体分割支路网络进行删除，得到所述训练好的人体质量检测模型。
需要说明的是，本发明实施例提供的人体图像质量检测装置可以应用于可以进行人体图像质量检测的智能手机、电脑、服务器等设备。
本发明实施例提供的人体图像质量检测装置能够实现上述方法实施例中人体图像质量检测方法实现的各个过程，且可以达到相同的有益效果。为避免重复，这里不再赘述。
参见图4，图4是本发明实施例提供的一种电子设备的结构示意图，如图4所示，包括：存储器402、处理器401及存储在所述存储器402上并可在所述处理器401上运行的人体图像质量检测方法的计算机程序，其中：
处理器401用于调用存储器402存储的计算机程序，执行如下步骤：
获取待检测图像，所述待检测图像包括目标人体图像；
通过训练好的人体质量检测模型对的待检测图像进行人体图像质量检测，得到所述目标人体图像的质量检测结果；
其中，所述人体质量检测模型根据样本人体图像、与所述样本人体图像对应的图像质量标签以及与所述样本人体图像对应的人体分割标签进行训练得到。
可选的，在所述通过训练好的人体质量检测模型对的待检测图像进行人体图像质量检测，得到所述目标人体图像的质量检测结果之前，处理器401执行的所述方法还包括：
获取样本图像以及初始人体质量检测模型；
对所述样本图像进行图像分割，得到样本人体图像，并根据分割结果对所述样本人体图像进行标注，得到所述样本人体图像对应的人体分割标签；
对所述样本图像进行图像质量标注，得到对应的图像质量标签；
基于所述样本人体图像、所述人体分割标签以及所述图像质量标签，构建得到训练数据集；
根据所述训练数据集对所述初始人体质量检测模型进行训练，得到所述训练好的人体质量检测模型。
可选的，所述样本图像包括多个样本人体图像，处理器401执行的所述对所述样本图像进行图像分割，得到样本人体图像，并根据分割结果对所述样本人体图像进行标注，得到所述样本人体图像对应的人体分割标签，包括：
对所述样本图像进行图像分割，得到第一样本人体图像和第二样本人体图像，所述第一样本人体图像为在所述样本图像中居中且面积最大的人体图像；
对所述第一样本人体图像和所述第二样本人体图像进行分类标注，得到所述第一样本人体图像和所述第二样本人体图像对应的人体分割标签。
可选的，处理器401执行的所述对所述样本图像进行图像质量标注，得到对应的图像质量标签，包括：
通过预设的图像处理方法对所述第一样本人体图像进行图像处理，得到第三样本图像；
根据所述预设的图像处理方法，对所述第三样本图像进行第一图像质量标注；
计算所述第二样本人体图像与所述第一样本人体图像之间的面积比；
根据所述面积比，对所述样本图像进行第二图像质量标注；
基于所述第一图像质量标注与所述第二图像质量标注，得到所述样本图像对应的图像质量标签。
可选的，处理器401执行的所述通过预设的图像处理方法对所述第一样本人体图像进行图像处理，得到第三样本图像，包括：
根据预设的截断方向和截断比例对所述第一样本人体图像进行截断处理，得到截断人体图像，所述截断方向和所述截断比例还用于对所述截断人体图像进行截断图像质量标注；
根据预设的遮挡方向和遮挡比例对所述第一样本人体图像进行遮挡处理，得到遮挡人体图像，所述遮挡方向和所述遮挡比例还用于对所述遮挡人体图像进行遮挡图像质量标注；
根据预设的模糊参数对所述第一样本人体图像进行模糊处理，得到模糊人体图像，所述模糊参数还用于对所述模糊人体图像进行模糊图像质量标注；
基于所述截断人体图像、所述遮挡人体图像以及所述模糊人体图像，得到第三样本图像。
可选的，所述初始人体质量检测模型包括主干网络、人体分割支路网络以 及人体质量支路网络，处理器401执行的所述根据所述训练数据集对所述初始人体质量检测模型进行训练，得到所述训练好的人体质量检测模型，包括：
通过所述主干网络提取所述样本图像的共用特征；
通过所述人体分割支路网络对所述共用特征进行预测，得到人体分割预测，以及通过所述人体质量支路网络对所述共用特征进行预测，得到图像质量预测；
计算所述人体分割预测与所述人体分割标签之间的第一损失，以及计算所述图像质量预测与所述图像质量标签之间的第二损失；
根据所述第一损失与所述第二损失，对所述初始人体质量检测模型进行参数的迭代调整，得到所述训练好的人体质量检测模型。
可选的，处理器401执行的所述根据所述第一损失与所述第二损失，对所述初始人体质量检测模型进行参数的迭代调整，得到所述训练好的人体质量检测模型，包括：
根据所述第一损失与所述第二损失，通过随机梯度下降调整所述主干网络、所述人体分割支路网络以及所述人体质量支路网络对应的网络参数；
当所述第一损失与所述第二损失最小，或者迭代次数达到预设次数时，停止训练，并将训练好的所述人体分割支路网络进行删除，得到所述训练好的人体质量检测模型。
本发明实施例提供的电子设备能够实现上述方法实施例中人体图像质量检测方法实现的各个过程，且可以达到相同的有益效果。为避免重复，这里不再赘述。
本发明实施例还提供一种计算机可读存储介质，计算机可读存储介质上存储有计算机程序，该计算机程序被处理器执行时实现本发明实施例提供的人体图像质量检测方法或应用端人体图像质量检测方法的各个过程，且能达到相同的技术效果，为避免重复，这里不再赘述。
本领域普通技术人员可以理解实现上述实施例方法中的全部或部分流程，是可以通过计算机程序来指令相关的硬件来完成，所述的程序可存储于一计算机可读取存储介质中，该程序在执行时，可包括如上述各方法的实施例的流程。其中，所述的存储介质可为磁碟、光盘、只读存储记忆体(Read-Only Memory，ROM)或随机存取存储器(Random Access Memory，简称RAM)等。
以上所揭露的仅为本发明较佳实施例而已，当然不能以此来限定本发明之权利范围，因此依本发明权利要求所作的等同变化，仍属本发明所涵盖的范围。

Claims (10)

  1. 一种人体图像质量检测方法，其特征在于，包括以下步骤：
    获取待检测图像，所述待检测图像包括目标人体图像；
    通过训练好的人体质量检测模型对所述待检测图像进行人体图像质量检测，得到所述目标人体图像的质量检测结果；
    其中，所述人体质量检测模型根据样本人体图像、与所述样本人体图像对应的图像质量标签以及与所述样本人体图像对应的人体分割标签进行训练得到。
  2. 如权利要求1所述的人体图像质量检测方法，其特征在于，在所述通过训练好的人体质量检测模型对所述待检测图像进行人体图像质量检测，得到所述目标人体图像的质量检测结果之前，所述方法还包括：
    获取样本图像以及初始人体质量检测模型；
    对所述样本图像进行图像分割，得到样本人体图像，并根据分割结果对所述样本人体图像进行标注，得到所述样本人体图像对应的人体分割标签；
    对所述样本图像进行图像质量标注，得到对应的图像质量标签；
    基于所述样本人体图像、所述人体分割标签以及所述图像质量标签，构建得到训练数据集；
    根据所述训练数据集对所述初始人体质量检测模型进行训练，得到所述训练好的人体质量检测模型。
  3. 如权利要求2所述的人体图像质量检测方法，其特征在于，所述样本图像包括多个样本人体图像，所述对所述样本图像进行图像分割，得到样本人体图像，并根据分割结果对所述样本人体图像进行标注，得到所述样本人体图像对应的人体分割标签，包括：
    对所述样本图像进行图像分割，得到第一样本人体图像和第二样本人体图像，所述第一样本人体图像为在所述样本图像中居中且面积最大的人体图像；
    对所述第一样本人体图像和所述第二样本人体图像进行分类标注，得到所述第一样本人体图像和所述第二样本人体图像对应的人体分割标签。
  4. 如权利要求3所述的人体图像质量检测方法，其特征在于，所述对所述样本图像进行图像质量标注，得到对应的图像质量标签，包括：
    通过预设的图像处理方法对所述第一样本人体图像进行图像处理，得到第三样本图像；
    根据所述预设的图像处理方法，对所述第三样本图像进行第一图像质量标注；
    计算所述第二样本人体图像与所述第一样本人体图像之间的面积比；
    根据所述面积比，对所述样本图像进行第二图像质量标注；
    基于所述第一图像质量标注与所述第二图像质量标注，得到所述样本图像对应的图像质量标签。
  5. 如权利要求4所述的人体图像质量检测方法，其特征在于，所述通过预 设的图像处理方法对所述第一样本人体图像进行图像处理，得到第三样本图像，包括：
    根据预设的截断方向和截断比例对所述第一样本人体图像进行截断处理，得到截断人体图像，所述截断方向和所述截断比例还用于对所述截断人体图像进行截断图像质量标注；
    根据预设的遮挡方向和遮挡比例对所述第一样本人体图像进行遮挡处理，得到遮挡人体图像，所述遮挡方向和所述遮挡比例还用于对所述遮挡人体图像进行遮挡图像质量标注；
    根据预设的模糊参数对所述第一样本人体图像进行模糊处理，得到模糊人体图像，所述模糊参数还用于对所述模糊人体图像进行模糊图像质量标注；
    基于所述截断人体图像、所述遮挡人体图像以及所述模糊人体图像，得到第三样本图像。
  6. 如权利要求2至5中任一所述的人体图像质量检测方法，其特征在于，所述初始人体质量检测模型包括主干网络、人体分割支路网络以及人体质量支路网络，所述根据所述训练数据集对所述初始人体质量检测模型进行训练，得到所述训练好的人体质量检测模型，包括：
    通过所述主干网络提取所述样本图像的共用特征；
    通过所述人体分割支路网络对所述共用特征进行预测，得到人体分割预测，以及通过所述人体质量支路网络对所述共用特征进行预测，得到图像质量预测；
    计算所述人体分割预测与所述人体分割标签之间的第一损失，以及计算所述图像质量预测与所述图像质量标签之间的第二损失；
    根据所述第一损失与所述第二损失，对所述初始人体质量检测模型进行参数的迭代调整，得到所述训练好的人体质量检测模型。
  7. 如权利要求6所述的人体图像质量检测方法，其特征在于，所述根据所述第一损失与所述第二损失，对所述初始人体质量检测模型进行参数的迭代调整，得到所述训练好的人体质量检测模型，包括：
    根据所述第一损失与所述第二损失，通过随机梯度下降调整所述主干网络、所述人体分割支路网络以及所述人体质量支路网络对应的网络参数；
    当所述第一损失与所述第二损失最小，或者迭代次数达到预设次数时，停止训练，并将训练好的所述人体分割支路网络进行删除，得到所述训练好的人体质量检测模型。
  8. 一种人体图像质量检测装置，其特征在于，所述装置包括：
    第一获取模块，用于获取待检测图像，所述待检测图像包括目标人体图像；
    检测模块，用于通过训练好的人体质量检测模型对的待检测图像进行人体图像质量检测，得到所述目标人体图像的质量检测结果；
    其中，所述人体质量检测模型根据样本人体图像、与所述样本人体图像对应的图像质量标签以及与所述样本人体图像对应的人体分割标签进行训练得到。
  9. 一种电子设备，其特征在于，包括：存储器、处理器及存储在所述存储器上并可在所述处理器上运行的计算机程序，所述处理器执行所述计算机程序时实现如权利要求1至7中任一项所述的人体图像质量检测方法中的步骤。
  10. 一种计算机可读存储介质，其特征在于，所述计算机可读存储介质上存储有计算机程序，所述计算机程序被处理器执行时实现如权利要求1至7中任一项所述的人体图像质量检测方法中的步骤。
PCT/CN2022/141457 2022-07-12 2022-12-23 人体图像质量检测方法、装置、电子设备及存储介质 WO2024011853A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202210819601.X 2022-07-12
CN202210819601.XA CN115311680A (zh) 2022-07-12 2022-07-12 人体图像质量检测方法、装置、电子设备及存储介质

Publications (1)

Publication Number Publication Date
WO2024011853A1 true WO2024011853A1 (zh) 2024-01-18

Family

ID=83857664

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/141457 WO2024011853A1 (zh) 2022-07-12 2022-12-23 人体图像质量检测方法、装置、电子设备及存储介质

Country Status (2)

Country Link
CN (1) CN115311680A (zh)
WO (1) WO2024011853A1 (zh)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115311680A (zh) * 2022-07-12 2022-11-08 青岛云天励飞科技有限公司 人体图像质量检测方法、装置、电子设备及存储介质

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112633255A (zh) * 2020-06-18 2021-04-09 上海高德威智能交通系统有限公司 目标检测方法、装置及设备
CN113011345A (zh) * 2021-03-11 2021-06-22 百度在线网络技术（北京）有限公司 图像质量检测方法、装置、电子设备、可读存储介质
CN113158860A (zh) * 2021-04-12 2021-07-23 烽火通信科技股份有限公司 基于深度学习多维度输出人脸质量评估方法与电子设备
CN113962965A (zh) * 2021-10-26 2022-01-21 腾讯科技（深圳）有限公司 图像质量评价方法、装置、设备以及存储介质
CN115311680A (zh) * 2022-07-12 2022-11-08 青岛云天励飞科技有限公司 人体图像质量检测方法、装置、电子设备及存储介质

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112633255A (zh) * 2020-06-18 2021-04-09 上海高德威智能交通系统有限公司 目标检测方法、装置及设备
CN113011345A (zh) * 2021-03-11 2021-06-22 百度在线网络技术（北京）有限公司 图像质量检测方法、装置、电子设备、可读存储介质
CN113158860A (zh) * 2021-04-12 2021-07-23 烽火通信科技股份有限公司 基于深度学习多维度输出人脸质量评估方法与电子设备
CN113962965A (zh) * 2021-10-26 2022-01-21 腾讯科技（深圳）有限公司 图像质量评价方法、装置、设备以及存储介质
CN115311680A (zh) * 2022-07-12 2022-11-08 青岛云天励飞科技有限公司 人体图像质量检测方法、装置、电子设备及存储介质

Also Published As

Publication number Publication date
CN115311680A (zh) 2022-11-08

Similar Documents

Publication Publication Date Title
US11200424B2 (en) Space-time memory network for locating target object in video content
CN108182394B (zh) 卷积神经网络的训练方法、人脸识别方法及装置
WO2022002150A1 (zh) 一种视觉点云地图的构建方法、装置
JP7208408B2 (ja) 検出モデルのトレーニング方法、装置、コンピュータデバイス及びコンピュータプログラム
WO2020215672A1 (zh) 医学图像病灶检测定位方法、装置、设备及存储介质
CN108764133B (zh) 图像识别方法、装置及系统
WO2019120115A1 (zh) 人脸识别的方法、装置及计算机装置
WO2020228525A1 (zh) 地点识别及其模型训练的方法和装置以及电子设备
WO2020228446A1 (zh) 模型训练方法、装置、终端及存储介质
WO2018108129A1 (zh) 用于识别物体类别的方法及装置、电子设备
WO2020107847A1 (zh) 基于骨骼点的跌倒检测方法及其跌倒检测装置
CN108549852B (zh) 基于深度网络增强的特定场景下行人检测器自动学习方法
CN114746898A (zh) 用于生成图像抠图的三分图的方法和系统
WO2019056503A1 (zh) 门店监控评价方法、装置及存储介质
WO2020252917A1 (zh) 一种模糊人脸图像识别方法、装置、终端设备及介质
CN110570435A (zh) 用于对车辆损伤图像进行损伤分割的方法及装置
WO2024011853A1 (zh) 人体图像质量检测方法、装置、电子设备及存储介质
CN110942456B (zh) 篡改图像检测方法、装置、设备及存储介质
CN113255557A (zh) 一种基于深度学习的视频人群情绪分析方法及系统
CN113297956A (zh) 一种基于视觉的手势识别方法及系统
CN111292333B (zh) 用于分割图像的方法和装置
TW202226054A (zh) 物件辨識裝置及物件辨識方法
CN112580581A (zh) 目标检测方法、装置及电子设备
WO2024011873A1 (zh) 目标检测方法、装置、电子设备及存储介质
CN113470093B (zh) 基于航拍图像处理的视频果冻效应检测方法、装置及设备

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22950960

Country of ref document: EP

Kind code of ref document: A1