A clothing key point detection method based on an optimized heat map supervision mechanism

Abstract

According to the clothing key point detection method based on the optimized heat map supervision mechanism provided in the embodiment, the clothing key point detection training set is calculated to obtain three scales of predicted heat maps through multi-resolution heat maps. Since the backbone network adopted is HRNet, high resolution can be maintained in the information transmission process. Then, by adding a Gaussian kernel with gradually reduced standard deviation in the gradually focusing heat map supervision mechanism, redundant spatial information can be brought while increasing the spatial resolution. Finally, the horizontal loss function and the mean square error are calculated, and the clothing key point detection network is trained in the reverse propagation manner to obtain the clothing key point detection network. By training the clothing key point detection network through the horizontal loss function, a better detection model can be trained without damaging the calculation performance too much, and the accuracy of the clothing key point detection is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to a clothing key point detection method based on an optimized heat map supervision mechanism. BACKGROUND

[0002] The purpose of clothing key point detection is to predict the position of key points defined on clothes, such as the corners of the collar, hem and sleeve. Detecting the position of key points as an underlying ring in clothing information recognition technology plays an important role in clothing position calibration, attribute prediction and other tasks.

[0003] The key point detection task originated from human pose estimation and face key point positioning. Traditional key point detection tasks are based on the method of regressing coordinates, extracting features from input images, and regressing key point coordinates from the output layer. However, the regression method has problems such as key point coupling in prediction, and the prediction performance is poor. The current proposed framework mainly includes deep fashion alignment (DFA) composed of three-stage convolutional neural networks and deep landmark network (DLAN) which combines selective dilated convolution and recursive spatial transformation. Both of these two frameworks are based on the method of directly regressing key point coordinates for prediction. In addition, there are methods that achieve more accurate landmark prediction performance through a global-local embedding module and a robust model Match R-CNN based on Mask R-CNN by establishing a large benchmark dataset DeepFashion2, which also achieves good detection effect.

[0004] The latest advanced research uses a hierarchical method to utilize structure diagram reasoning method to strengthen the structural layout relationship between key points. This method defines different parts of key points as nodes of a graph and clusters features in the idea of graph neural network. Although the use of complex information transmission method and reasoning module increases the model capacity, it also brings increased computational complexity.

[0005] In summary, in the task of clothing key point detection, most advanced methods are based on the classic heat map supervision method to fully design the feature extraction process. Few methods conduct in-depth research on the heat map supervision method in the prediction link. SUMMARY

[0006] To solve the above problems, a clothing key point detection method based on an optimized heat map supervision mechanism is provided. The present application adopts the following technical solutions:

[0007] The application provides a clothing key point detection method based on an optimized heat map supervision mechanism, and has the characteristics that the method comprises the following steps: S1-1, an input image is input into a pre-trained clothing key point detection network to obtain an output heat map of different resolutions for the position of each key point; S1-2, the output heat map is processed by difference upsampling and arithmetic averaging to obtain a final output heat map; and S1-3, the final output heat map is output, wherein the training process of the clothing key point detection network comprises the following steps: S2-1, an original data set comprising an original image, a clothing bounding box and key points is pre-processed to obtain image features, and the image features are subjected to data enhancement to obtain a clothing key point detection training set comprising processed images; S2-2, a multi-resolution heat map supervision module in a to-be-trained clothing key point detection network is used to obtain predicted heat maps of three scales for the clothing key point detection training set; S2-3, a step-by-step focusing heat supervision module is used to calculate a real heat map for the clothing key point detection training set; S2-4, a horizontal loss function and a mean square error are calculated by comparing the predicted heat map and the real heat map; S2-5, a total loss function is calculated according to the horizontal loss function and the mean square error; and S2-6, the clothing key point detection network is trained by back propagation in a predetermined gradient rotation method according to the mean square error and the total loss function, so that the clothing key point detection network is obtained.

[0008] The clothing key point detection method based on the optimized heat map supervision mechanism also has the following technical features: the horizontal loss function is calculated as follows: In the formula, HL represents the horizontal loss function, y and are pixel values on the real heat map and the predicted heat map respectively, is the maximum confidence value in the model predicted heat map. The mean square error is calculated as follows: In the formula, MSE represents the mean square error, y and are pixel values on the real heat map and the predicted heat map respectively.

[0009] The clothing key point detection method based on the optimized heat map supervision mechanism also has the following technical features: the total loss function is calculated by the mean square error and the horizontal loss function, and the calculation process is as follows: L total =MSE+α×HL In the formula, L total is the total loss function, MSE is the mean square error, HL is the horizontal loss function, and α is a hyperparameter.

[0010] The clothes key point detection method based on the optimized heat map supervision mechanism also has the following technical features: the multi-resolution heat map supervision module includes an upsampling unit, a three-path feature convolution unit, and a multi-size supervision unit; the upsampling unit is configured to obtain scale feature maps at three different resolutions by twice deconvolution upsampling of a processed image; the three-path feature convolution unit is configured to obtain output heat maps at three different scales by twice convolution of the scale feature maps at the three different resolutions through a 1x1 convolution kernel and by adjusting the number of channels to match the number of predicted key points; and the multi-size supervision unit is configured to supervise the output heat maps at the three different scales for a single-channel heat map at the three scales.

[0011] The clothes key point detection method based on the optimized heat map supervision mechanism also has the following technical features: the step-by-step focusing heat supervision module generates a real heat map by performing Gaussian blur on the positions of key points in an original image; the step-by-step focusing heat supervision module includes a Gaussian kernel, and the Gaussian kernel has a standard deviation that gradually decreases with the advancement of training time and decreases to a small value by reducing to half of the previous standard deviation.

[0012] The clothes key point detection method based on the optimized heat map supervision mechanism also has the following technical features: the predetermined gradient rotation method is configured to use only mean square error to perform gradient backpropagation on a to-be-trained clothes key point detection network in the first five training cycles, use a total loss function to perform gradient backpropagation on the to-be-trained clothes key point detection network and update the parameters of the network model, use a larger standard deviation in the early stage of training to ensure the convergence speed of the model, and gradually reduce the standard deviation to a small value as the training cycle progresses.

[0013] The clothes key point detection method based on the optimized heat map supervision mechanism also has the following technical features: the data enhancement is configured to process image features by random flipping, random rotation, and random cropping to obtain a clothes key point detection data set; the random flipping is configured to flip the training data image in the vertical direction in a clockwise or counterclockwise direction with a predetermined probability; the random rotation is configured to randomly select an angle from [-30°, +30°] to rotate the training data image; and the random cropping is configured to randomly crop the area outside the clothes bounding box in the training data image.

[0014] The clothes key point detection method based on the optimized heat map supervision mechanism provided by the application can also have the following technical features: the image preprocessing includes the following steps: step S2-1-1, obtaining a single-channel heat map according to the key points by calculation; step S2-1-2, scaling the resolution size of the original image; step S2-1-3, obtaining image features by normalizing the single-channel heat map and the scaled input image.

[0015] Inventive action and effect

[0016] According to the clothes key point detection method based on the optimized heat map supervision mechanism, the clothes key point detection training set is used to calculate three scale prediction heat maps through multi-resolution heat map calculation. Since the backbone network used is HRNet, high resolution can be maintained during information transmission. Then, by adding a Gaussian kernel with gradually decreasing standard deviation in the gradually focusing heat map supervision mechanism, redundant spatial information can be brought while increasing the spatial resolution. Finally, the clothes key point detection network is obtained by calculating the horizontal loss function, mean square error, and training the clothes key point detection network in the form of back propagation. Training the clothes key point detection network through the horizontal loss function can train a better detection model without significantly affecting the calculation performance, thereby improving the accuracy of clothes key point detection. By adding a Gaussian kernel with gradually decreasing standard deviation in the gradually focusing heat map supervision mechanism, redundant spatial information can be brought while increasing the spatial resolution. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 is a flowchart of the clothes key point detection network in the embodiment of the application;

[0018] Figure 2 is a schematic diagram of the clothes key point detection network in the embodiment of the application;

[0019] Figure 3 is a flowchart of the clothes key point detection network training process in the embodiment of the application;

[0020] Figure 4 is a schematic diagram of the backbone network in the embodiment of the application;

[0021] Figure 5 is a schematic diagram of the multi-resolution supervision module in the embodiment of the application;

[0022] Figure 6 is the standard deviation change value of the gradually focusing heat map supervision mechanism in the embodiment of the application;

[0023] Figure 7 is a schematic diagram of the horizontal loss function in the embodiment of the application; and

[0024] Figure 8 is the effect diagram of the clothing key point detection result in the embodiment of the application. DETAILED DESCRIPTION

[0025] In order to make the technical means, creative features, purposes and effects realized by the application easy to understand, the following will specifically describe a clothing key point detection method based on an optimized heat map supervision mechanism in combination with embodiments and drawings.

[0026] <EMBODIMENT>

[0027] In the embodiment, the input image and the key points are passed through the clothing key point detection network to obtain the final output heat map capable of accurately positioning the key points.

[0028] Figure 1 is the flowchart of the clothing key point detection network in the embodiment of the application.

[0029] Figure 2 is the schematic diagram of the clothing key point detection network in the embodiment of the application.

[0030] As shown in Figure 1 and Figure 2 , the process of the clothing key point detection network includes steps S1-1 to S1-3.

[0031] Step S1-1, the input image is passed through the pre-trained clothing key point detection network to calculate the output heat map of different resolutions for the position of each key point.

[0032] Step S1-2, the output heat map is processed by difference upsampling and arithmetic average to obtain the final output heat map.

[0033] In the embodiment, the output heat map is of three different resolutions, and the output heat maps of the three different resolutions are 128*128*8 pixels, 256*256*8 pixels and 512*512*8 pixels. The output heat maps of resolutions of 128*128*8 pixels and 256*256*8 pixels are upsampled to have a resolution of 512*512*8 pixels, and the output heat map of resolution of 512*512*8 pixels is processed by arithmetic average to obtain the final output heat map of 512*512*8 pixels.

[0034] Step S1-3, output the final output heat map (model final detection result).

[0035] The position coordinate with the maximum confidence value in the heat map corresponds to the position coordinate of the key point. The key points are the coordinates of the collar, the cuff, the waist, and the hem in the clothing image, which not only indicate the functional areas of the clothing, but also implicitly capture the bounding box of the clothing, so as to better distinguish the design, pattern, and category of the clothing.

[0036] Figure 3 FIG. 1 is a flowchart of a clothing key point detection network training process in an embodiment of the present application.

[0037] As shown in FIG. 2, the training process of the clothing key point detection network specifically includes the following steps S2-1 to S2-6. Figure 3

[0038] Step S2-1, image preprocessing is performed on the original data set prepared in advance including the original image, the clothing bounding box, and the key point, and the obtained image features are obtained through data enhancement to obtain a clothing key point detection training set including the processed image.

[0039] In this embodiment, the original data set is a DeepFashion-C data set. The DeepFashion-C data set contains a total of 289222 clothing images, and each image corresponds to the position of the clothing bounding box and the position of the key point.

[0040] The coordinates of the key points are derived from the annotation of the DeepFashion-C data set.

[0041] The image preprocessing includes the following steps S2-1-1 to S2-1-3.

[0042] Step S2-1-1, a single-channel heat map is obtained by calculating according to the key points.

[0043] Step S2-1-2, the resolution size of the input image is scaled.

[0044] In this embodiment, the resolution size of the input image is scaled to 512x512 pixels.

[0045] Step S2-1-3, the single-channel heat map and the scaled input image are processed by normalization to obtain image features.

[0046] In this embodiment, the image preprocessing uses the HRNet network as the backbone network, which contains 4 parallel subnetworks gradually increasing from high resolution to low resolution, and detects the input data with a resolution size of 3x512x512.

[0047] Figure 4 FIG. 3 is a schematic diagram of a backbone network in an embodiment of the present application.

[0048] As​Figure 4 As shown, the backbone network contains parallel subnets that upsample and downsample the feature map blocks through convolution units. The feature map is the feature information at each stage in the network operation process, the convolution unit is a convolution calculation that only increases the depth of the feature map block without changing the size of the feature map block, the downsample is a convolution calculation that reduces the size of the feature map block, and the upsample is an interpolation calculation that increases the size of the feature map block.

[0049] The four parallel subnets are respectively a one-stage parallel subnet, a two-stage parallel subnet, a three-stage parallel subnet, and a four-stage parallel subnet.

[0050] The one-stage parallel subnet only has a single branch of 128x128 resolution feature map; the two-stage parallel subnet has two double branches of 128x128 resolution feature map and 64x64 resolution feature map; the three-stage parallel subnet has three triple branches of 128x128 resolution feature map, 64x64 resolution feature map, and 32x32 resolution feature map; and the four-stage parallel subnet has four quadruple branches of 128x128 resolution feature map, 64x64 resolution feature map, 32x32 resolution feature map, and 16x16 resolution feature map. In each stage, there are a large number of inter-branch information connections, transmissions, and calculations. Finally, the 128x128 resolution feature map, the 64x64 resolution feature map, the 32x32 resolution feature map, and the 16x16 resolution feature map are upsampled to the same resolution of 128x128 and become a whole final feature map through concatenation. Since high resolution is maintained while low resolution is fully extracted, the final feature map has excellent low-level spatial information and high-level semantic information, which can provide a solid feature basis for subsequent prediction.

[0051] The network components in the HRNet mainly consist of multiple layers of convolution (Convolutional) layers, batch normalization (Batch Normalization, BN) layers, rectified linear unit (Rectified Linear Unit, ReLU), block operations, and tran operations.

[0052] The convolution layer extracts local features from the image through the convolution kernel in a sliding manner, and its parameters are adjusted iteratively with the training of the model to achieve automatic feature extraction.

[0053] The batch normalization layer normalizes the output of the intermediate layer of the model, thereby accelerating the convergence of the network model while maintaining the consistency of the distribution. The calculation formula of the output layer is as follows:

[0054]

[0055] In the formula, x is the input of the batch normalization layer, y is the output of the batch normalization layer, E[x] is the mean of x, Var[x] is the variance of x, and γ and β are preset learnable parameters.

[0056] The ReLU layer is used to increase the fitting ability of the model through nonlinear activation, and the calculation formula of the ReLU layer output is as follows:

[0057] f(x) = max(0, x)

[0058] In the formula, f(x) is the output of the ReLU layer, and x is the input of the batch normalization layer.

[0059] The block operation is a 3x3 convolution operation with a stride of 1 and padding of 1; the tran operation connects two parallel subnets of different stages. The channel number is changed through a 1x1 convolution between features of the same resolution; the channel number and resolution size are changed through a 3x3 convolution operation with a stride of 2 and followed by a BN layer and a ReLU layer between features of different resolutions; the fuse operation fuses feature blocks of different resolutions, reduces the resolution through a 3x3 convolution with a stride of 2, increases the resolution through interpolation upsampling, and then performs concat fusion and changes the channel number through a 1x1 convolution, and a BN layer and a ReLU layer are followed after each convolution.

[0060] The image features obtained through image preprocessing are subjected to data enhancement.

[0061] The data enhancement processes the preprocessed image through random flipping, random rotation, and random cropping to obtain image features.

[0062] The random flipping is to flip the training data image in the vertical direction in a clockwise or counterclockwise direction with a preset probability.

[0063] The random rotation is to randomly select an angle from [-30°, +30°] to rotate the training data image.

[0064] The random cropping is to randomly crop the area outside the clothing bounding box in the training data image.

[0065] In step S2-2, the multi-resolution heat map supervision module in the to-be-trained clothing key point detection network is used to obtain three scales of predicted heat maps for the clothing key point detection training set.

[0066] Figure 5 is a schematic diagram of the multi-resolution supervision module in the embodiment of the application.

[0067] As shown in Figure 5 , the multi-resolution heat map supervision module includes an upsampling unit, a three-path feature convolution unit, and a multi-size supervision unit.

[0068] The up-sampling unit is configured to up-sample the processed image through twice deconvolution to obtain scale feature maps at three different resolutions (1, 2 and 3 in the figure, respectively).

[0069] The three-path feature convolution unit is configured to perform twice convolution on the three scale feature maps through a 1*1 convolution kernel, and obtain predicted heat maps at three different scales (4, 5 and 6 in the figure, respectively) by adjusting the number of channels to match the number of predicted key points.

[0070] The multi-size supervision unit is configured to supervise the output feature maps at three scales at small, medium and large sizes for the single-channel heat maps (7, 8 and 9 in the figure, respectively).

[0071] In step S2-3, the real heat map is calculated by the step-by-step focusing heat supervision module for the clothing key point detection training set.

[0072] The step-by-step focusing heat supervision module generates the real heat map by performing Gaussian blur on the positions of the key points in the original image.

[0073] The step-by-step focusing heat supervision module includes a Gaussian kernel.

[0074] The Gaussian kernel has a standard deviation that gradually decreases as the training time progresses, and the standard deviation is reduced to a small value by reducing it to half of the previous standard deviation.

[0075] Figure 6 is the standard deviation change value of the gradually focusing heat map supervision mechanism in the embodiment of the present application.

[0076] As shown in Figure 6 , the Gaussian kernel uses a larger standard deviation in the first five cycles of training, and the standard deviation of the Gaussian kernel gradually decreases in the later period.

[0077] In the first five cycles of training, the mean square error is used to perform gradient backpropagation on the clothing key point detection network to be trained, so as to ensure the convergence speed of the model; and in the later period, the total loss function is used to perform gradient backpropagation on the clothing key point detection network to be trained and update the parameters of the network model.

[0078] In step S2-4, the horizontal loss function and the mean square error are calculated by comparing the predicted heat map and the real heat map.

[0079] Figure 7 is a schematic diagram of the horizontal loss function in the embodiment of the present application.

[0080] As shown in Figure 7As shown, the maximum confidence value in the predicted heat map is calculated by calculating the height difference and the horizontal difference.

[0081] The horizontal loss function is calculated as follows:

[0082]

[0083] In the formula, HL represents the horizontal loss function, y and are pixel values on the real heat map and the predicted heat map, respectively. is the maximum confidence value in the predicted heat map.

[0084] The mean square error is calculated as follows:

[0085]

[0086] In the formula, MSE represents the mean square error, y and are pixel values on the real heat map and the predicted heat map, respectively.

[0087] In step S2-5, the total loss function is calculated by the horizontal loss function and the mean square error.

[0088] The total loss function is calculated by the mean square error and the horizontal loss function, and the calculation process is as follows:

[0089] L total = MSE + a x HL

[0090] In the formula, L total represents the total loss function, MSE represents the mean square error, HL represents the horizontal loss function, and a is a hyperparameter.

[0091] In this embodiment, the hyperparameter a in the total loss function is 0.5.

[0092] In step S2-6, the back propagation is performed by the mean square error and the total loss function in a predetermined gradient rotation method, so as to train the clothing key point detection network to obtain the clothing key point detection network.

[0093] The predetermined gradient rotation method is that, in the first five cycles of training, only the mean square error is used to perform gradient back propagation on the clothing key point detection network to be trained, and then the total loss function is used to perform gradient back propagation on the clothing key point detection network to be trained and update the parameters of the network model.

[0094] A larger standard deviation is used in the early stage of training to ensure the convergence speed of the model, and then the standard deviation is gradually reduced to a smaller value as the training cycle gradually proceeds.

[0095] In this embodiment, the training is performed on the backbone network to be trained by an Adam optimizer, where the learning rate is set to 0.00011, the learning rate decay coefficient is 0.9, the batch size is 16, and the training period is 30.

[0096] Figure 8 is an effect diagram of the clothing key point detection result in the embodiment of the application.

[0097] As shown in Figure 8 , Figure 7 each key point in each input image can be accurately identified and labeled, and it can be seen that the clothing key point detection model of this embodiment can improve the accuracy of clothing key point detection when performing the key point detection task on the public dataset DeepFashion-C, without excessively damaging the computing performance, so that the average detection normalized error is reduced from 0.0441 to 0.0334, thereby reaching the accuracy of the first-line detection method in the academic field.

[0098] Effects of the embodiment

[0099] According to the clothing key point detection method based on the optimized heat map supervision mechanism provided in this embodiment, different resolution output heat maps are calculated for the position of each key point in the input image through the pre-trained clothing key point detection network, and the position of the key point is provided as important bottom space information source, which has a wide range of application places; and the final output heat map is obtained through difference upsampling and arithmetic average processing, thereby realizing effective detection of clothing key points. The training process of the clothing key point detection network is as follows: the clothing key point detection training set is calculated to obtain three scales of predicted heat maps through multi-resolution heat map calculation, and since the backbone network used is HRNet which has better generalization, it can maintain high resolution during information transmission; then, by adding a Gaussian kernel with gradually decreasing standard deviation in the gradually focusing heat map supervision mechanism, redundant spatial information can be brought in while increasing the spatial resolution; finally, the horizontal loss function and the mean square error are calculated, and the clothing key point detection network is trained in the form of back propagation to obtain the clothing key point detection network. Therefore, through the clothing key point detection network of the application, better detection effect can be achieved without excessively damaging the computing performance, thereby improving the accuracy of clothing key point detection.

[0100] In the embodiment, 1+2y is used as the index, so that the position with a larger value of y can also obtain a larger gradient for back propagation, so that the gradient of the backbone network is related to the distance between the predicted value and the true value, and the back propagation gradient can obtain the maximum value at the confidence peak value of the true heat map (i.e. the key point coordinate position).

[0101] In the embodiments, when data enhancement is performed, considering the clothing bounding box, random clipping is avoided, so that the clothing is not cut off, and the phenomenon of inaccurate detection results caused by cutting off the clothing is avoided.

[0102] The above embodiments are only used to illustrate the specific implementation manners of the present application, and the present application is not limited to the description range of the above embodiments.

Claims

1. A clothing key point detection method based on an optimized heat map supervision mechanism, characterized in that, Comprise: Step S1-1, the input image is calculated by the pre-trained clothing key point detection network for each key point position. The output heat map of different resolutions is obtained; Step S1-2, the output heat map is obtained by difference upsampling and arithmetic average processing to obtain the final output heat map; Step S1-3, output the final output heat map, Wherein, the training process of the clothing key point detection network is specifically: Step S2-1, the original image, clothing bounding box and key point of the original data set prepared in advance are preprocessed, and the image features obtained are enhanced by data to obtain a clothing key point detection training set including processed images; Step S2-2, for the clothing key point detection training set, the multi-resolution heat map supervision module in the to-be-trained clothing key point detection network is used to obtain three scales of predicted heat map; Step S2-3, for the clothing key point detection training set, the real heat map is calculated by the step-by-step focusing heat supervision module; Step S2-4, the horizontal loss function and the mean square error are calculated by comparing the predicted heat map and the real heat map; Step S2-5, the total loss function is calculated according to the horizontal loss function and the mean square error; Step S2-6, the mean square error and the total loss function are used to perform back propagation by the predetermined gradient rotation method, so as to train the clothing key point detection network and obtain the clothing key point detection network.

2. The clothing key point detection method based on the optimized heat map supervision mechanism according to claim 1, wherein: wherein The horizontal loss function calculation process is as follows: wherein HL is the horizontal loss function, y and are pixel values on the real heat map and the predicted heat map, respectively, is the maximum confidence value in the predicted heat map, The calculation process of the mean square error is as follows: where MSE is the mean square error, y and are pixel values on the real heat map and the predicted heat map, respectively.

3. The clothing key point detection method based on the optimized heat map supervision mechanism according to claim 2, wherein: wherein, The total loss function is calculated by the mean square error and the horizontal loss function, and the calculation process is as follows: L total = MSE + a x HL In the formula, L total is the total loss function, MSE is the mean square error, HL is the horizontal loss function, and a is a hyperparameter.

4. The clothing key point detection method based on the optimized heat map supervision mechanism according to claim 1, wherein: wherein, The multi-resolution heat map supervision module comprises an upsampling part, a three-way feature convolution part and a multi-size supervision part, The upsampling part is to obtain three scale feature maps under three different resolutions by twice deconvolution upsampling of the processed image, The three-way feature convolution part is to obtain output heat maps under three different scales by twice convolution of the scale feature maps under three different resolutions through 1x1 convolution kernel, and to match the number of predicted key points by adjusting the number of channels, The multi-size supervision part is to supervise the output heat maps under three scales at three scales for single-channel heat maps.

5. The clothing key point detection method based on the optimized heat map supervision mechanism according to claim 1, wherein: wherein The step-by-step focusing heat supervision module generates a real heat map by Gaussian blur on the position of the key point in the original image, The step-by-step focusing heat supervision module includes a Gaussian kernel, The Gaussian kernel has a standard deviation that gradually decreases with the advancement of the training time, and the standard deviation is reduced to a small value in a manner of reducing to half of the previous standard deviation.

6. The clothing key point detection method based on an optimized heatmap supervision mechanism according to claim 1, characterized in that: wherein The predetermined gradient rotation method is to use the mean square error to perform gradient backpropagation on the clothing key point detection network to be trained only in the first five training cycles, to perform gradient backpropagation on the clothing key point detection network to be trained through the total loss function, and to update the parameters of the network model, The standard deviation is used to ensure the convergence speed of the model at the beginning of training, and the standard deviation is gradually reduced to a small value as the training cycle progresses.

7. The clothing key point detection method based on an optimized heatmap supervision mechanism according to claim 1, characterized in that: wherein The data augmentation processes the image features by means of random flipping, random rotation, and random cropping to obtain the clothing key point detection data set, The random flipping is to flip the training data image in the vertical direction in a clockwise or counterclockwise direction with a predetermined probability, The random rotation is to randomly select an angle from [-30°, +30°] to rotate the training data image, The random cropping is to randomly crop the area outside the clothing bounding box in the training data image.

8. The clothing key point detection method based on an optimized heatmap supervision mechanism according to claim 1, characterized in that: wherein The image preprocessing includes the following steps: Step S2-1-1, obtaining a single-channel heatmap by calculation according to the key points; Step S2-1-2, scaling the resolution size of the original image; Step S2-1-3, obtaining the image features by normalizing the single-channel heatmap and the scaled input image.

Images