Low-resolution human body posture estimation method and device based on posture distillation and medium

Abstract

The invention discloses a low-resolution human body posture estimation method and device based on posture distillation and a medium. The method comprises the steps that paired human body posture estimation data sets are constructed; obtaining a pre-trained model after pre-training and fine tuning, and taking the pre-trained model as a teacher model in posture distillation; constructing a posture distiller; constructing a low-resolution human body posture detection model; on the paired high and low resolution training sets, using a posture distiller to decompose human body posture features obtained by the teacher model, and guiding the low resolution network to learn different levels of human body posture features in stages; and performing human body posture estimation and prediction by using the trained low-resolution human body posture estimation model to obtain a human body posture estimation and prediction result. According to the posture distiller provided by the invention, the human body posture characteristics of the teacher model are decomposed on different scales and transmitted to the student model, so that the detection capability of the student model in a low-resolution environment is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of human posture estimation, and in particular to a low-resolution human posture estimation method, device and storage medium based on posture distillation. Background Art

[0002] Human pose estimation is an important branch of computer vision, which aims to detect the joint position, orientation and posture of the human body from images or videos. This technology plays an important role in application scenarios such as human action recognition, behavior analysis and human-computer interaction. In recent years, human pose estimation methods based on deep learning have emerged continuously, significantly improving the accuracy of prediction.

[0003] However, when the image resolution is reduced, human pose estimation faces great challenges. Due to the lack of detailed information in low-resolution images, the human structure and key points in the image become blurred, making it difficult for existing models to extract effective human pose features. This leads to the poor performance of existing human pose estimation network models on low-resolution images, which seriously affects the detection performance of the algorithm.

[0004] In view of the difficulty of extracting human posture features from low-resolution images, existing methods mainly adopt two strategies: one is to use multi-scale feature fusion technology to enhance the expressiveness of human posture features by combining feature maps of different resolutions; the other is to use feature super-resolution technology to restore the detailed information of human posture features by upsampling low-resolution feature maps. However, these methods also have some problems, such as high computational overhead, high computational complexity, and excessive processing that introduces noise, which reduces the practicality of the network model and affects the prediction accuracy of the network model. Summary of the invention

[0005] In order to improve the practicality and detection ability of low-resolution network models and solve the challenges faced by low-resolution human posture estimation, this paper proposes a low-resolution human posture estimation method based on posture distillation. The core idea of ​​this method is to design a new posture distiller, which can decompose the human posture features of the teacher model at different scales and pass them to the student model, thereby improving the posture perception ability of the student model on low-resolution images, and then realizing accurate low-resolution human posture detection.

[0006] To achieve the above object, the present invention adopts the following technical solutions: The method comprises the following steps: S1, construct a paired human pose estimation dataset, including high-resolution and low-resolution paired training sets, low-resolution validation sets, and test sets; S2, obtain the pre-trained and fine-tuned pre-trained model as the teacher model in pose distillation; S3, construct the posture distiller, including the spatial alignment layer, contour decoupling layer, and posture encoding layer; S4, constructing a low-resolution human posture detection model, including several types of neural network computing layers, parameters of each neural network computing layer, and connection methods between each neural network computing layer; S5, on paired high- and low-resolution training sets, uses a posture distiller to decompose the human posture features of the teacher model at different scales and pass them to the student model, so that the student model can learn the global and local features of human posture at different levels, thereby improving its prediction accuracy on low-resolution images; S6, using the trained low-resolution human posture estimation model to perform human posture estimation prediction to obtain a human posture estimation prediction result.

[0007] Furthermore, step S1 includes the following steps: S1.1, constructing a high-resolution and low-resolution paired training set, wherein the training set is based on the COCO dataset, which contains the annotation information of the key points of the human body. Affine transformations of different scales are performed on the training images in the dataset to obtain the corresponding high-resolution and low-resolution images and annotations, and then data enhancement operations such as rotation, blurring, and translation are performed on the high-resolution and low-resolution images to improve the generalization ability of the model and form the final training set; S1.2, construct a low-resolution image verification set, which is based on the COCO dataset. The verification images in the dataset are transformed with affine transformations of different scales to obtain low-resolution images for verifying the performance and effect of the model. S1.3, construct a low-resolution image test set, which is based on the COCO dataset. The test images in the dataset are transformed with affine transformations of different scales to obtain low-resolution images for testing the generalization ability and robustness of the model.

[0008] Further, step S2 includes the following steps: The HRNet-W32 network is selected as the backbone network of the teacher model. The network can maintain high-resolution feature representation throughout the process, and improve the fusion of semantic information and position information through multi-resolution sub-networks and information interaction, thereby obtaining more detailed human posture features; The network is fine-tuned on the data-augmented high-resolution COCO dataset to improve the generalization of features extracted by the teacher model.

[0009] Compared with the existing distillation methods, the posture distiller proposed in the present invention can encode complete and relatively pure human posture features from a high-resolution teacher network, reducing the error and noise of the teacher network; at the same time, it maintains small model parameters, reduces computational complexity and memory consumption; and can provide phased guidance for the student network, realizing layer-by-layer learning from shallow to deep, significantly improving the student model's ability to perceive human posture in extreme environments.

[0010] Further, step S3 includes the following steps: S3.1, constructing a spatial alignment layer, which is composed of multiple deconvolution units, convolution units, and RELU functions, and is used to reduce the spatial dimension of high-resolution features while retaining information such as the spatial position of the features; S3.2, constructing a contour decoupling layer, wherein the contour decoupling layer includes a global nested edge detection layer and a multi-layer perceptron, wherein the former is used to extract human contour information from features, and the latter is used to denoise and optimize the extracted human contour features to reduce interference from background noise; S3.3, construct a posture encoding layer, which is composed of multiple linear convolution units and nonlinear convolution units stacked together. The linear encoding unit is composed of multiple 1x1 convolutions and normalization operations to extract human posture features. The nonlinear operation generates multiple human posture features of different scales by multiple 1x1 convolutions, RELU activation functions and normalization operations. For an image, after the high-resolution teacher model, high-resolution human posture features can be obtained, effectively eliminating background interference. Since the high-resolution human posture features still have a small spatial scale, this will affect the subsequent human contour detection and extraction effect. To this end, step S3.1 is required to spatially align the features, and step S3.1 includes the following steps: Through the deconvolution unit in the spatial alignment layer, the features obtained by the teacher model are spatially amplified by 4 times, and the number of channels is halved, thereby increasing the spatial scale of the features and preparing for subsequent contour extraction; Then, the features are reorganized by the convolutional units in the spatial alignment layer to enrich the spatial position and global information of the features and remove the influence of the noise channel.

[0011] Furthermore, the contour decoupling layer decouples the reorganized features to extract pure and complete human body contour features. Step S3.2 includes the following steps: First, high-resolution image features are input into the VGG16 backbone model of HED to extract features of different scales and depths.

[0012] Then, the feature map at the end of each stage is sent to the corresponding hidden layer for multi-scale feature extraction, thereby obtaining multiple human body contour prediction maps with richer spatial structure information, improving the adaptability and robustness of the model. Then, in order to maintain global consistency, the feature maps of all hidden layers are fused together to obtain a fused feature map; Next, in order to eliminate the differences between different feature maps, a weighted average operation is performed on all feature maps to generate the final human body contour prediction map. Finally, a multi-layer perceptron is used to denoise and optimize the human body contour prediction image to reduce the impact of background noise.

[0013] Furthermore, the posture coding layer performs coding optimization on the human body contour prediction map to generate human body posture features that can be used to guide the student model. Step S3.3 includes the following steps: The posture encoding layer uses a linear transformation matrix to perform low-dimensional feature mapping on each pixel of the human body contour prediction image, thereby acquiring and enhancing the human body posture features; Then, the nonlinear convolution unit is used to perform nonlinear transformation on the human posture features to generate posture features of different scales.

[0014] Next, a lightweight low-resolution human pose estimation model is constructed as the student model. Step 4 includes the following steps: Select a suitable backbone network model, which can be any existing convolutional neural network model, such as the ResNetHRNet-W32 convolutional neural network model; According to the characteristics of the backbone network, the number of downsampling is reduced to adapt to low-resolution images.

[0015] Furthermore, the student model and some parameters of the posture distiller are trained on the high- and low-resolution paired training sets. During the training process, the characteristics of the teacher model output are defined as , the output features of the student model ,The training process uses KL divergence to guide the student model learning, and the guidance process is: ,in, is a hyperparameter that controls the soft signal output of the teacher network and the student network. The specific step 5 includes the following steps: A pair of high- and low-resolution images are randomly selected from the training set and input into the teacher model and student model for feature extraction respectively; The pose distiller decouples multi-scale pose features from the features of the teacher model and uses large-scale features to guide the shallow perception of the student model; At the same time, small-scale features are used to guide the deep perception of the student model, thereby comprehensively improving the posture detection ability of the student model in a low-resolution environment.

[0016] Finally, the MSELOSS loss function is used to measure the gap between the prediction results and the annotations, and the parameters of the model are optimized to obtain a trained low-resolution human pose estimation model. MSE is defined as: in, is the sample size, is the true value, is the model prediction value.

[0017] Further, step S6 includes the following steps: Randomly select an input sample from the test set and obtain the input data after data preprocessing; The input data is a trained low-resolution human pose estimation model, which outputs the coordinates of key points; According to the requirements of the data set, the key points are connected into a visual result graph.

[0018] On the other hand, the present invention further discloses a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the above method.

[0019] On the other hand, the present invention further discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the above method.

[0020] It can be seen from the above technical solution that in order to effectively solve the problem of difficulty in extracting human posture features on low-resolution images and improve the practicality of the network model. The present invention proposes a low-resolution human posture estimation method based on posture distillation, which uses a high-resolution dataset and model as a teacher and a low-resolution dataset and model as a student. The posture distiller proposed in the present invention decomposes the human posture features of the teacher model at different scales and transfers them to the student model, thereby improving the detection ability of the student model in a low-resolution environment. In addition, the student model based on posture distillation does not require too many training parameters and can be deployed to a wider range of devices.

[0021] Specifically, the present invention discloses a low-resolution human posture estimation method based on posture distillation, and its steps include: constructing a paired human posture estimation data set; obtaining a pre-trained model after pre-training and fine-tuning as a teacher model in posture distillation; constructing a posture distiller; constructing a low-resolution human posture detection model; on paired high- and low-resolution training sets, using the posture distiller to decompose the human posture features obtained by the teacher model, and guiding the low-resolution network to learn human posture features at different levels in stages; using the trained low-resolution human posture estimation model to perform human posture estimation prediction, and obtain human posture estimation prediction results.

[0022] Compared with the prior art, the present invention has the following beneficial effects: The present invention proposes a low-resolution human posture estimation method based on posture distillation to solve the problem that the existing model is difficult to extract effective posture features in a low-resolution environment, and greatly improves the detection capability of the model; The present invention designs a new posture distiller, which can extract complete and pure human posture features from the teacher model, reduce the interference of the environment and background, and can improve the posture perception ability of the student model in stages, significantly improving the detection performance of the student model; The posture distiller and student model of the present invention have small parameters, low calculation amount, small memory usage, are suitable for training and deployment on a variety of devices, and realize human posture detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a workflow diagram of a low-resolution human posture estimation method model based on posture distillation provided by an embodiment of the present invention; Figure 2 It is a structural diagram of a posture distiller provided by an embodiment of the present invention; Figure 3 It is a structural diagram of a spatial alignment layer in a posture distiller provided by an embodiment of the present invention; Figure 4 It is a structural diagram of a profile decoupling layer in a posture distiller provided by an embodiment of the present invention; Figure 5 It is a structural diagram of a posture encoding layer in a posture distiller provided by an embodiment of the present invention; Figure 6 It is a training flow chart of a low-resolution human posture estimation method based on posture distillation provided by an embodiment of the present invention; Figure 7 , Figure 8 They are schematic diagrams of low-resolution human posture estimation results based on the student model. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments.

[0025] like Figure 1As shown, an embodiment of the present invention proposes a low-resolution human posture estimation method based on posture distillation. The method uses a high-resolution dataset and a model as a teacher, and a low-resolution dataset and a model as a student. The human posture features of the teacher model are decomposed at different scales and passed to the student model through the posture distiller proposed by the present invention. This method overcomes the challenge of difficult extraction of human posture features on low-resolution images, improves the practicality of the model, and realizes accurate low-resolution human posture prediction. The specific steps are as follows: S1, construct a paired human pose estimation dataset, including high-resolution and low-resolution paired training sets, low-resolution validation sets, and test sets; S2, obtain the pre-trained and fine-tuned pre-trained model as the teacher model in pose distillation; S3, construct the posture distiller, including the spatial alignment layer, contour decoupling layer, and posture encoding layer; S4, constructing a low-resolution human posture detection model, including several types of neural network computing layers, parameters of each neural network computing layer, and connection methods between each neural network computing layer; S5, on paired high- and low-resolution training sets, uses a posture distiller to decompose the human posture features of the teacher model at different scales and pass them to the student model, so that the student model can learn the global and local features of human posture at different levels, thereby improving its prediction accuracy on low-resolution images; S6, using the trained low-resolution human posture estimation model to perform human posture estimation prediction to obtain a human posture estimation prediction result.

[0026] Specifically, step S1 includes the following steps: S1.1, constructing a high-resolution and low-resolution paired training set, wherein the training set is based on the COCO dataset, which contains the annotation information of the coordinates of the key points of the human body. The training images in the dataset are subjected to affine transformations of different scales to obtain the corresponding high-resolution and low-resolution images and annotations, and then the high-resolution and low-resolution images are subjected to data enhancement operations such as rotation, blurring, and translation to improve the generalization ability of the model and form the final training set; S1.2, construct a low-resolution image verification set, which is based on the COCO dataset. The verification images in the dataset are transformed with affine transformations of different scales to obtain low-resolution images for verifying the performance and effect of the model. S1.3, construct a low-resolution image test set, which is based on the COCO dataset. The test images in the dataset are transformed with affine transformations of different scales to obtain low-resolution images for testing the generalization ability and robustness of the model.

[0027] Step S2 includes the following steps: The HRNet-W32 network is selected as the backbone network of the teacher model. The network can maintain high-resolution feature representation throughout the process, and improve the fusion of semantic information and position information through multi-resolution sub-networks and information interaction, thereby obtaining more detailed human posture features; The network is fine-tuned on the data-augmented high-resolution COCO dataset to improve the generalization of features extracted by the teacher model.

[0028] Compared with the existing distillation methods, the posture distiller proposed in the present invention can encode noise-free human posture features from a high-resolution teacher network, reducing the error and noise of the teacher network; at the same time, it maintains small model parameters, reduces computational complexity and memory consumption; it can provide staged guidance for the student network, realize layer-by-layer learning from shallow to deep, and significantly improve the student model's ability to perceive human posture in extreme environments.

[0029] like Figure 2 As shown, step S3 includes three feature extraction layers including a spatial alignment layer, a contour decoupling layer, and a posture encoding layer. The specific steps of each layer are as follows: S3.1, constructing a spatial alignment layer, which is composed of a deconvolution unit and a convolution unit, realizing spatial alignment of high-resolution teacher model features to low-resolution student model features, while retaining information such as the spatial position of the features; S3.2, constructing a contour decoupling layer, wherein the contour decoupling layer includes a global nested edge detection layer and a multi-layer perceptron, wherein the former is used to extract human contour information from features, and the latter is used to denoise and optimize the extracted human contour features to reduce interference from background noise; S3.3, construct a posture coding layer, which is composed of a plurality of linear coding units and nonlinear convolution units stacked together, uses linear coding to extract human posture features, and then generates multi-level human posture features through nonlinear transformation, and guides the student model to learn in turn.

[0030] For a high-resolution image , through the HRNet-W32 network, high-resolution human posture features can be obtained Due to the high-resolution human posture features It still has a small spatial scale, which will affect the subsequent human contour detection and extraction effect. To this end, step S3.1 is required to spatially align the features.

[0031] like Figure 3 As shown, step S3.1 includes the following steps: The high-resolution human posture features are transformed through the deconvolution unit in the spatial alignment layer. Magnify the space by 4 times and halve the number of channels, thereby increasing the spatial scale of the feature and preparing for the subsequent contour extraction operation; Then, through the convolution unit in the spatial alignment layer, the features are reorganized to obtain a feature with appropriate spatial scale and rich information. , which can enrich the spatial position and global information of the features and remove the influence of the noise channel. The expression is as follows:

[0032] in is the feature obtained by the spatial alignment layer, is the deconvolution unit in the spatial alignment layer, CONV is the convolutional unit in the spatial alignment layer.

[0033] like Figure 4 As shown, step S3.2 is used to decouple the reorganized features and extract complete and noise-free human body contour features. Step S3.2 includes the following steps: The global nested edge detection layer in the contour decoupling layer is used to reorganize the features First, the VGG16 backbone model is used to extract the reorganized features to obtain features of different scales, and then it is sent to the corresponding hidden layer for multi-scale feature extraction, so as to obtain multiple human body contour prediction images with richer spatial structure information. , thus avoiding the problem of single scale and level, the expression is as follows:

[0034] in, Refers to the predicted human body contour map, Refers to the global nested edge detection layer.

[0035] Then, a weighted fusion strategy is used to linearly superimpose all the human body contour prediction maps to obtain a more complete and detailed human body contour prediction map. , the expression is as follows:

[0036] in, Refers to the predicted human body contour map, It means that the obtained sub-features are spliced ​​and reorganized into a new feature. Indicates that the concatenation operation is fused along the channel dimension.

[0037] Finally, the multi-layer perceptron is used to denoise and optimize the human body contour prediction map to reduce the impact of background noise. The expression is as follows:

[0038] in, It is the human body contour prediction map obtained after optimization in the multi-layer perceptron, MLP

[0039] like Figure 5 As shown, the posture encoding layer performs encoding optimization on the human body contour prediction map to generate human body posture features that can be used to guide the student model. Step S3.3 includes the following steps: The posture encoding layer uses the linear transformation matrix to predict the human body contour Perform low-dimensional feature mapping on each pixel to obtain and enhance human posture features; Then, the nonlinear convolution unit is used to perform multi-scale transformation on the human posture features to generate posture features at different levels. , the expression is as follows:

[0040] in, is the posture feature obtained after the posture encoding layer is encoded, LCONV

[0041] Next, a lightweight low-resolution human pose estimation model is constructed as a student model. Further, step 4 includes the following steps: Select a suitable backbone network model, which can be any existing convolutional neural network model, such as the SimpleBaseline model convolutional neural network model; According to the characteristics of the backbone network, the number of downsampling is appropriately reduced to adapt to low-resolution images.

[0042] like Figure 6 As shown, the student model and some parameters in the posture distiller are trained on the high-resolution and low-resolution paired training sets. Step 5 includes the following steps: Randomly select a pair of high-resolution images from the training set and low-resolution images , respectively input into the teacher model to obtain high-resolution features and the student model to obtain shallow low-resolution features ; The pose distiller disentangles multi-scale human pose features from the features of the teacher model. , and then use KL divergence in the shallow layer to guide the student model to learn. The expression is as follows:

[0043] in, is a hyperparameter that controls the soft signal output of the teacher network and the student network.

[0044] At the same time, small-scale features are used to guide the learning of the student model in depth, thereby comprehensively improving the posture ability of the student model in a low-resolution environment. The expression is as follows:

[0045] in Represents the deep features of the student model.

[0046] Finally, the MSELOSS loss function is used to measure the gap between the prediction results and the annotations, and the parameters of the model are optimized to obtain a trained low-resolution human pose estimation model. MSE is defined as: in, is the sample size, is the true value, is the model prediction value.

[0047] Step S6 includes the following steps: Randomly select an input sample from the test set and obtain the input data after data preprocessing; The input data is a trained low-resolution human pose estimation model, which outputs the coordinates of key points; According to the requirements of the data set, the key points are connected into a visual result graph.

[0048] Table 1

[0049] Table 1 shows the average prediction accuracy comparison of the SimpleBaseline model based on pose distillation and the original SimpleBaseline model on the low-resolution validation set. It can be clearly seen from the table that the SimpleBaseline model based on pose distillation has a significant improvement in accuracy compared to the original SimpleBaseline model, showing higher precision. Figure 7 , Figure 8 The results of the student model on the low-resolution human posture estimation task are shown. It can be seen from the example that the low-resolution human posture estimation method provided by the embodiment of the present invention accurately locates the key points of various parts of the human body, effectively solves the problem that the model is difficult to extract effective human posture features in a low-resolution environment, and significantly improves the detection performance of low-resolution human posture estimation.

[0050] On the other hand, the present invention further discloses a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the above method.

[0051] On the other hand, the present invention further discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the above method.

[0052] In another embodiment provided in the present application, a computer program product comprising instructions is also provided, which, when executed on a computer, enables the computer to execute any low-resolution human posture estimation method based on posture distillation in the above-mentioned embodiments.

[0053] It is understandable that the system, device and storage medium provided in the embodiments of the present invention correspond to the method provided in the embodiments of the present invention, and the explanation, examples and beneficial effects of the relevant contents can refer to the corresponding parts in the above methods.

[0054] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from a website site, a computer, a server or a data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website site, computer, server or data center. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server or a data center that includes one or more available media integrated. The available medium may be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive Solid State Disk (SSD)), etc.

[0055] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0056] Each embodiment in this specification is described in a related manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0057] The above embodiments 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. However, these 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.

Claims

1. A low-resolution human posture estimation method based on posture distillation, characterized in that: The steps include: S1, construct a paired human pose estimation dataset, including high-resolution and low-resolution paired training sets, low-resolution validation sets, and test sets; S2, obtain the pre-trained and fine-tuned pre-trained model as the teacher model in pose distillation; S3, construct the posture distiller, including the spatial alignment layer, contour decoupling layer, and posture encoding layer; S4, constructing a low-resolution human posture detection model, namely a student model, including a neural network computing layer, parameters of each neural network computing layer, and a connection method between each neural network computing layer; S5, on paired high- and low-resolution training sets, uses a posture distiller to decompose the human posture features of the teacher model at different scales and pass them to the student model, so that the student model can learn the global and local features of human posture at different levels, thereby improving its prediction accuracy on low-resolution images; S6, using the trained low-resolution human posture estimation model to perform human posture estimation prediction to obtain a human posture estimation prediction result.

2. The low-resolution human posture estimation method based on posture distillation according to claim 1, characterized in that: Step S1 includes the following steps: S1.1, construct a high-resolution and low-resolution paired training set, the training set is based on the COCO dataset and contains the annotation information of the key points of the human body, the training images of the training set are subjected to affine transformation of different scales to obtain the corresponding high-resolution and low-resolution images and annotations, and then the high-resolution and low-resolution images are subjected to rotation, blurring, and translation data enhancement operations to improve the generalization ability of the model and form the final training set; S1.2, construct a low-resolution image verification set, which is based on the COCO dataset. The verification images in the verification set are transformed with affine at different scales to obtain low-resolution images for verifying the performance and effect of the model. S1.3, construct a low-resolution image test set. The test set is based on the COCO dataset. Affine transformations of different scales are performed on the test images in the test set to obtain low-resolution images for testing the generalization ability and robustness of the model.

3. The low-resolution human posture estimation method based on posture distillation according to claim 1, characterized in that: Step S2 includes the following steps: Select HRNet-W32 network as the backbone network of the teacher model; The network is fine-tuned on a data-augmented high-resolution training set to improve the generalization of features extracted by the teacher model.

4. The low-resolution human posture estimation method based on posture distillation according to claim 1, characterized in that: Step S3 includes the following steps: S3.1, constructing a spatial alignment layer, wherein the spatial alignment layer includes a plurality of deconvolution units, convolution units, and RELU functions, and is used to reduce the spatial dimension of high-resolution features while retaining the spatial position information of the features; S3.2, constructing a contour decoupling layer, wherein the contour decoupling layer includes a global nested edge detection layer and a multi-layer perceptron, wherein the global nested edge detection layer is used to extract human contour information from features, and the multi-layer perceptron is used to denoise and optimize the extracted human contour features to reduce interference from background noise; S3.3, construct a posture encoding layer, which is composed of a plurality of linear convolution units and a nonlinear convolution unit stacked together. The linear encoding unit is composed of a plurality of 1x1 convolutions and normalization operations for extracting human posture features, and the nonlinear unit is composed of a plurality of 1x1 convolutions, RELU activation functions and normalization operations for generating a plurality of human posture features of different scales.

5. The low-resolution human posture estimation method based on posture distillation according to claim 4, characterized in that: For a high-resolution image , through the HRNet-W32 network, high-resolution human posture features are obtained ; The features are spatially aligned in step S3.1, which specifically includes the following steps: The high-resolution human posture features are transformed through the deconvolution unit in the spatial alignment layer. Magnify the space by 4 times and halve the number of channels, thereby increasing the spatial scale of the feature and preparing for the subsequent contour extraction operation; Then, the convolutional unit in the spatial alignment layer reorganizes the human body contour and edge information to obtain a feature with appropriate spatial scale and rich information. , the expression is as follows: in is the feature obtained by the spatial alignment layer, is the deconvolution unit in the spatial alignment layer, CONV is the convolution unit in the spatial alignment layer; The features are extracted through the convolution unit, and then optimized and reorganized through the RELU activation function to obtain high-resolution detail features.

6. The low-resolution human posture estimation method based on posture distillation according to claim 4, characterized in that: Step S3.2 is used to decouple the reorganized features and extract complete and noise-free human body contour features. Step S3.2 specifically includes the following steps: The global nested edge detection layer in the contour decoupling layer is used to reorganize the features First, the VGG16 backbone model is used to extract the reorganized features to obtain features of different scales, and then it is sent to the corresponding hidden layer for multi-scale feature extraction, so as to obtain multiple human body contour prediction images with richer spatial structure information. , thus avoiding the problem of single scale and level, the expression is as follows: in, Refers to the predicted human body contour map, refers to the global nested edge detection layer; Then, a weighted fusion strategy is used to linearly superimpose all the human body contour prediction maps to obtain a more complete and detailed human body contour prediction map. , the expression is as follows: in, Refers to the predicted human body contour map, It means that the obtained sub-features are spliced ​​and reorganized into a new feature. Indicates that the concatenation operation is fused along the channel dimension; Finally, the multi-layer perceptron is used to denoise and optimize the human body contour prediction map to reduce the impact of background noise. The expression is as follows: in, It is the human body contour prediction map obtained after optimization in the multi-layer perceptron. MLP 7. The low-resolution human posture estimation method based on posture distillation according to claim 4, characterized in that: The posture encoding layer performs encoding optimization on the human body contour prediction map to generate human body posture features that can be used to guide the student model. Step S3.3 specifically includes the following steps: The posture encoding layer uses the linear transformation matrix to predict the human body contour Perform low-dimensional feature mapping on each pixel to obtain and enhance human posture features; Then, the nonlinear convolution unit is used to perform multi-scale transformation on the human posture features to generate posture features at different levels. , the expression is as follows: in, is the posture feature obtained after encoding by the posture encoding layer, LCONV 8. The low-resolution human posture estimation method based on posture distillation according to claim 1, characterized in that: Step S5 includes the following steps: Define the characteristics of the teacher model output as , the output features of the student model ,The training process uses KL divergence to guide the student model learning, and the guidance process is: ,in, It is a hyperparameter that controls the soft signal output of the teacher network and the student network; The specific step S5 includes the following steps: A pair of high- and low-resolution images are randomly selected in the training set, and features are extracted using the teacher model and the student model respectively; The pose distiller decouples multi-scale pose features from the features of the teacher model and uses large-scale features to guide the shallow perception of the student model; At the same time, small-scale features are used to guide the deep perception of the student model, improving the prediction performance and robustness of the student model at low resolution; Finally, the MSELOSS loss function is used to measure the gap between the prediction results and the annotations, and the parameters of the model are optimized to obtain a trained low-resolution human pose estimation model. MSE is defined as: in, is the sample size, is the true value, is the model prediction value.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the processor is caused to perform the steps of the method according to any one of claims 1 to 8.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the computer program is executed by the processor, the processor is caused to perform the steps of the method according to any one of claims 1 to 8.

Images