A two-stage method for photographic body position recognition and body surface anatomical landmark detection based on viewpoint assistance

Abstract

The present invention discloses a two-stage method for photographic position recognition and surface anatomical landmark detection based on view-assisted perspective, comprising the following steps: S1: regrouping 23 photographic positions to generate view-assisted perspective labels; S2: using the network model pre-trained based on the ImageNet data set as the initialization weight to train the network for the position classification stage; S3: using the network model pre-trained based on the ImageNet data set as the initialization weight to train the network for the surface anatomical landmark detection stage; S4: inputting the data to be tested in the data set into the optimized two-stage network; S5: decoding two groups of heat maps into coordinates, and using the average Gaussian similarity to evaluate the two groups of coordinates; S6: using the high detection performance landmark point type sequence as prior information, selecting and processing the two groups of heat maps, and outputting the coordinate results. The present invention has an improved level of automated processing, which can greatly reduce the workload and infection risk of radiology technicians.

Description

Technical Field

[0001] The invention relates to the field of digital image processing, and in particular to a two-stage method of photographic body position recognition and body surface anatomical landmark detection based on view angle assistance. Background Art

[0002] When radiology technicians perform X-ray photography, after confirming the shooting position, they need to guide the examinee to make the correct photography position and align the X-ray detector with the anatomical landmarks on the body surface of the shooting site. In order to reduce the workload and infection risk of technicians, it is particularly important to realize auxiliary X-ray photography through sensors and computers. The most critical part of auxiliary X-ray photography research is to automatically identify the photography position and detect the anatomical landmarks on the body surface in the visible light image.

[0003] In the current research on auxiliary X-ray photography, there is only one dataset called X-ray photography position recognition and surface anatomical landmark detection and a research on photography position recognition and surface anatomical landmark detection methods based on multi-task learning. In the multi-task learning method, during training, it is necessary to judge whether the optimal model needs to be updated after this training based on the performance evaluation results of the main branch. This method can only ensure that the performance of the main branch is optimal, but other branches may not achieve the best performance. As another branch, photography position recognition often cannot achieve the best classification performance in the multi-task learning method. However, the classification results of the photography position provide prior information for the detection of surface anatomical landmarks, so the wrong photography position classification will cause all the surface anatomical landmarks under the photography position to be detected incorrectly. In summary, the photography position classification branch that cannot achieve the highest classification performance limits the performance of surface anatomical landmark detection. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a two-stage method of photographic position recognition and surface anatomical landmark detection based on viewpoint assistance, which solves the problem that multi-task branches of multi-task learning methods cannot achieve the best performance at the same time, and accurately realizes photographic position recognition and surface anatomical landmark detection of auxiliary X-ray photography.

[0005] To achieve the above object, the present invention adopts the following technical solution.

[0006] A two-stage method of photographic body position recognition and body surface anatomical landmark detection based on view angle assistance includes the following steps:

[0007] S1: Regroup the 23 photography positions and generate viewpoint labels;

[0008] S2: Based on the network model pre-trained with the ImageNet dataset as the initial weight, the photographic position classification phase network including the backbone network, the photographic position branch, the view classification branch, and the body surface anatomical landmark detection is trained;

[0009] S3: Based on the network model pre-trained with the ImageNet dataset as the initial weight, the network for body surface anatomical landmark detection is trained;

[0010] S4: Input the test data in the data set into the optimized two-stage network to obtain the photographic position classification results and the body surface anatomical landmark heat map;

[0011] S5: Decode the two sets of heat maps into coordinates, use the average Gaussian similarity to evaluate the two sets of coordinates, and obtain a sequence of landmark types with high detection performance;

[0012] S6: Using the high detection performance landmark point type sequence as prior information, the two sets of heat maps obtained in step S4 are selected and processed, and the coordinate results are output.

[0013] Furthermore, in step S1, the 23 photographic postures are analyzed and regrouped according to the viewing angle of the human body in the image, that is, the relative angle between the human body direction and the camera.

[0014] Furthermore, in step S2, the height or width of the human body bounding box is expanded to a fixed aspect ratio of height: width = 4:3, the bounding box is cropped from the image, and then adjusted to a fixed size of 384 pixels × 288 pixels as the input image, and the image is input to the network in the photography posture stage. The backbone network uses HRNetV2-W48 to extract feature maps N of four resolutions 41 ∈R N*ω*H / 4*W / 4 , N 42 ∈R N*2ω*H / 8*W / 8 , N 43 ∈R N*4ω*H / 16*W / 16 , N 44 ∈R N*8ω*H / 32*W / 32 , w is set to 48, H and W are the height and width of the input image.

[0015] Furthermore, the structure of the view classification branch in step S2 is as follows: first, N 41 , N 42 , N 43 , N 44 The convolutional layers are fed into the bottleneck layer of ResNet to increase their channels to 128, 256, 512, and 1024, respectively. The high-resolution feature maps are then downsampled using a 3×3 convolution with 256 channels and a step size of 2, and then added to the second-highest resolution feature maps to fuse them into low-level features. This process is repeated three times to obtain a minimum resolution feature map with 1024 channels. The number of feature map channels is then increased to 2048 through a 1×1 convolutional layer. Finally, the feature maps are sequentially passed through an average pooling layer and a fully connected layer to obtain the results of view 6 classification.

[0016] Furthermore, the photographic position classification branch in step S2 is the main branch of the photographic position classification stage, and adopts a network structure similar to that of the view classification branch, except that an SE module with a compression factor of 8 is added to perform feature recalibration on feature maps of different resolutions. First, the SE module is inserted into the feature map with the highest resolution whose channels are increased to 128, and the channel of the highest resolution feature map is adjusted; secondly, the SE module is added to the secondary resolution feature map obtained by adding and fusion of the high-resolution feature map after downsampling the high-resolution feature map and the secondary resolution feature map, and then recalibration is performed after the fusion of multiple resolution features; finally, the SE module is added to the lowest resolution feature map with a channel number of 2048 obtained by 1×1 convolution, and the last channel feature recalibration before classification is performed; finally, the recalibrated feature map passes through the average pooling layer and the fully connected layer in turn to obtain the result of photographic position 23 classification.

[0017] Furthermore, the loss of the view classification branch in step S2 is L poscls , using cross entropy loss; the loss of the photography position branch is L angcls , using cross entropy loss; the loss of the anatomical landmark detection branch is L landmark , using mean square error loss; the loss in the training phase is set to:

[0018] L=0.00001*L poscls +0.00001*L angcls +L landmark (1)

[0019] Random rotation and random scaling are used for image data enhancement. The random rotation range is ±45 degrees, and the random scaling range is ±35%. Based on the idea of ​​transfer learning, the network model pre-trained on ImageNet is used for model migration on the training set in the dataset. The Adam optimizer is used for training optimization. The basic learning rate is set to 1e-3, and it is reduced to 1e-4 and 1e-5 in the 70th and 100th epochs respectively. The training process ends within 110 epochs.

[0020] Furthermore, the stage of detecting body surface anatomical landmarks in step S3 includes the following steps:

[0021] The height or width of the human body bounding box is expanded to a fixed aspect ratio of height: width = 4:3. The bounding box is cropped from the image and then resized to a fixed size of 384 pixels × 288 pixels as the input image. The image is input to the network in the photography posture stage. First, HRNetV2-W48 is used as the backbone network to extract feature maps N of four resolutions. 41 , N 42 , N43 , N 44 ; Then N 41 Predict the 80-channel heatmap using L through 1×1 convolution landmark As the loss function of this stage, the data augmentation method, pre-training model and optimizer selection in the training stage are the same as those in the photographic position classification stage.

[0022] Furthermore, the body surface anatomical landmark heat map in step S4 includes the body surface anatomical landmark heat map output by the network in the photographic position classification stage in step S2. Figure 1 and the body surface anatomical landmark heat map 2 output by the network in the body surface anatomical landmark detection phase in step S3.

[0023] Furthermore, in step S5, the two sets of coordinates are evaluated using the average Gaussian similarity, which is expressed as follows:

[0024]

[0025] Where i is the index of the landmark type; d j is the Euclidean distance between the predicted coordinates of the jth landmark of the i-th type of anatomical landmark (each landmark will only appear once in each human instance) and the corresponding true coordinates; s is the scale factor, which is the size of the bounding box of the human instance where the jth i-type key point is located; κ i is a constant that controls the attenuation rate of the i-th type of anatomical surface landmark; N is the total number of i-type landmark points in the test image.

[0026] Furthermore, the selection processing and coordinate extraction in step S6 includes the following steps: using the photographic position classification result of the photographic position classification stage to filter the body surface anatomical landmark heat map 2, if the serial numbers corresponding to the channels that are not 0 in the filtered body surface anatomical landmark heat map 2 contain a high detection performance landmark point type sequence, then replacing the channels in the body surface anatomical landmark heat map 2 corresponding to these serial numbers with the body surface anatomical landmark heat map corresponding to these serial numbers Figure 1 Finally, the processed body surface anatomical landmark heat map 2 is processed according to formula (2) to obtain the final detection coordinates.

[0027] Compared with the prior art, the beneficial effects of the present invention are: 1) Two networks are designed to decouple the two tasks of photographic position recognition and surface anatomical landmarks, solving the problem that the multi-task branches in the multi-task learning method cannot achieve the best performance at the same time. 2) 97.40% photographic position classification accuracy and 80.1AP surface anatomical landmark detection rate are achieved, which is the method that achieves the highest photographic position classification performance and surface anatomical landmark detection performance. 3) The level of automated processing is improved, providing a core image processing method for the research of auxiliary X-ray photography, which can greatly reduce the workload and infection risk of radiology technicians. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 The figure is an overall flow chart of the two-stage method of the present invention based on perspective-assisted photographic body position recognition and body surface anatomical landmark detection. DETAILED DESCRIPTION

[0029] The basic idea of ​​the present invention is to decouple the two tasks and train two networks to perform photographic position recognition and surface anatomical landmark detection in steps, namely the photographic position classification stage and the surface anatomical landmark detection stage, so as to simultaneously explore the highest performance of photographic position recognition and surface anatomical landmark detection. Then, a selection processing module is designed. The module uses the evaluation result of the average Gaussian similarity between the predicted value and the true value of each surface anatomical landmark as prior information, and uses the two groups of surface anatomical landmark outputs generated in the two stages for selection processing. The method achieves the best detection effect on all types of landmarks. Figure 1 As shown, a two-stage method of photographic body position recognition and body surface anatomical landmark detection based on viewpoint assistance includes the following steps:

[0030] Step S1: The X-ray body position recognition and surface anatomical landmark detection dataset contains 23 body positions and 80 surface anatomical landmarks. The 23 body positions serve as the supervisory information for the body position classification branch. The 80 surface anatomical landmarks are converted into 80-channel heat maps as the supervisory information for the surface anatomical landmark detection branch. Among them, the anatomical landmark with coordinates marked as (0,0) will generate a heat map channel with all 0 values.

[0031] By analyzing 23 types of photographic positions, the 23 types of photographic positions are regrouped according to the perspective of the human body in the image, that is, the relative angle between the human body and the camera, so that each image in the data set generates another label without re-labeling. Specifically, photographic position category 1 is an image taken from the front by the camera when the human body makes an anatomical posture. That is, photographic position category 1 is the human body direction parallel to the Z axis of the camera coordinate system, which is defined here as a perspective of 0°. Taking photographic position category 1 as a reference, five groups are obtained according to the range of the perspective, that is, the range of the absolute value of the angle between the human body direction and the Z axis of the camera coordinate system. These five groups are: perspective of 0°, perspective of (0,90°), perspective of 90°, perspective of (90,180°), and perspective of 180°. The labels corresponding to the perspectives of the above five groups are 1-5. In addition, photographic position categories 22 and 23 are images of human sitting postures collected, which are classified as category 6. The above-mentioned perspective labels with 6 categories are used as supervision information for the perspective classification branch.

[0032] Step S2: Based on the idea of ​​multi-task learning, in the stage of photographic position classification, two auxiliary tasks, body surface decoupling landmark detection and viewpoint classification, are set to enhance the performance of photographic position classification. The network in the stage of photographic position classification includes a backbone network, a photographic position classification branch, a viewpoint classification branch, and a body surface anatomical landmark detection branch.

[0033] The height or width of the human body bounding box is expanded to a fixed aspect ratio of height: width = 4:3, the bounding box is cropped from the image, and then resized to a fixed size (384×288) as the input image. The image is input to the network in the photography posture stage. First, HRNetV2-W48 is used as the backbone network to extract feature maps N of four resolutions. 41 ∈R N *ω*H / 4*W / 4 , N 42 ∈R N*2ω*H / 8*W / 8 , N 43 ∈R N*4ω*H / 16*W / 16 , N 44 ∈R N*8ω*H / 32*W / 32 , w is set to 48, H and W are the height and width of the input image.

[0034] The view classification branch first transforms N 41 , N 42 , N 43 , N 44 The channels are fed into the bottleneck layer of ResNet to increase to 128, 256, 512, and 1024, respectively. The high-resolution feature map is then downsampled using a 3×3 convolution with 256 channels and a stride of 2, and then added to the second-highest resolution feature map to fuse into low-level features. After repeating this process three times, a minimum resolution feature map with 1024 channels (H / 32×W / 32) is obtained. The number of feature map channels is then increased to 2048 through a 1×1 convolution layer. Finally, the present invention passes the feature map through an average pooling layer and a fully connected layer in sequence to obtain the result of view 6 classification.

[0035] The photographic position classification branch is the main branch of the photographic position classification stage. It adopts a network structure similar to the view classification branch. The difference is that an SE module with a compression factor of 8 is added to recalibrate the feature maps of different resolutions. First, the SE module is inserted into the feature map with the highest resolution whose channels are increased to 128, and the channel of the highest resolution feature map is adjusted. Secondly, the SE module is added to the secondary resolution feature map obtained by adding and fusing the high-resolution feature map after downsampling the high-resolution feature map and the secondary resolution feature map, and then recalibrated after the fusion of multiple resolution features. Finally, the SE module is added to the lowest resolution feature map with a channel number of 2048 obtained by 1×1 convolution, and the last channel feature recalibration before classification is performed. Finally, the present invention passes the feature map through the average pooling layer and the fully connected layer in sequence to obtain the results of 23 photographic position classification.

[0036] N 41 The data is input into the body surface anatomical landmark detection branch and a 80-channel heat map is predicted through a 1×1 convolution.

[0037] The loss of the view classification branch is L poscls , using cross entropy loss; the loss of the photography position branch is L angcls , using cross entropy loss; the loss of the anatomical landmark detection branch is L landmark , using mean square error loss. The loss in the training phase is set to:

[0038] L=0.00001*L poscls +0.00001*L angcls +L landmardk (1)

[0039] Random rotation and random scaling are used for image data augmentation. The random rotation range is ±45 degrees, and the random scaling range is ±35%.

[0040] Based on the idea of ​​transfer learning, the network model pre-trained on ImageNet is used for model migration on the training set in the dataset, and the Adam optimizer is used for training optimization. The basic learning rate is set to 1e-3, and it is reduced to 1e-4 and 1e-5 in the 70th and 100th epochs respectively. The training process ends within 110 epochs.

[0041] Step S3: Surface anatomical landmark detection stage only detects surface anatomical landmarks and uses only the heat map formed by the coordinate encoding of the 80 surface anatomical landmarks in the data set for training.

[0042] The height or width of the human body bounding box is expanded to a fixed aspect ratio of height: width = 4:3, the bounding box is cropped from the image, and then resized to a fixed size (384 pixels × 288 pixels) as the input image. The image is input to the network in the photography posture stage, and HRNetV2-W48 is first used as the backbone network to extract feature maps N of four resolutions. 41 , N 42 , N 43 , N 44 Then N 41 Predict the 80-channel heatmap using L through 1×1 convolution landmark As the loss function of this stage, the data enhancement method, pre-training model and optimizer selection in the training stage are the same as those in the photography posture classification stage.

[0043] Step S4: Input the data to be tested in the data set into the network of the photographic position classification stage to obtain the photographic position classification result, the body surface anatomical landmark thermal Figure 1 The data to be tested in the data set is input into the network of the body surface anatomical landmark detection stage to obtain the body surface anatomical landmark heat map 2.

[0044] Step S5: According to the classification results generated in the photographic position classification stage, find the channel number corresponding to the surface anatomical landmark defined under the photographic position in the 80-channel heat map. According to the sequence number, the values ​​of the channels corresponding to the sequence numbers in the two sets of heat maps obtained in the two stages are kept unchanged, and the values ​​of the channels corresponding to the surface anatomical landmarks of other photographic position categories are converted to 0. Then, for the two processed sets of heat maps, determine the position (m) of the maximum value and the position (s) of the second largest value of each channel, and move a quarter of the length from the maximum activation to the second maximum activation to obtain the coordinates of each surface anatomical landmark corresponding to the two sets of heat maps:

[0045]

[0046] Then, the average Gaussian similarity (gs) between the predicted value and the true value of each anatomical landmark is used to evaluate the two sets of anatomical landmark coordinates:

[0047]

[0048] Where i is the index of the landmark type; d j is the Euclidean distance between the predicted coordinates of the jth landmark of the i-th type of anatomical landmark (each landmark will only appear once in each human instance) and the corresponding true coordinates; s is the scale factor, which is the size of the bounding box of the human instance where the jth i-type key point is located; κ iis a constant that controls the decay rate of the i-th type of anatomical landmark; N is the total number of landmarks of type i in the test image. It is worth noting that only when the photographic position is correctly classified, the landmarks in the image will be used to calculate the average Gaussian similarity. {gs} i It can directly represent the detection effect of the algorithm on the i-th anatomical landmark. i The detection effects of different detection algorithms on the same body surface anatomical landmarks can be compared. i The value ranges from 0 to 1. The closer it is to 1, the better the algorithm detects the anatomical landmarks of the i-th type.

[0049] Compare the evaluation results of the average Gaussian similarity of the two groups of body surface anatomical landmark coordinates, and output a sequence of landmark types whose detection effect of the first stage landmarks is better than that of the second stage landmarks. This sequence of landmark types with high detection performance is used as fixed prior information for subsequent tests.

[0050] Step S6: Use the photographic position classification results of the photographic position classification stage to filter the body surface anatomical landmark heat map 2. If the serial numbers corresponding to the channels in the filtered body surface anatomical landmark heat map 2 that are not 0 contain a high detection performance landmark point type sequence, then replace the channels in the body surface anatomical landmark heat map 2 corresponding to these serial numbers with the body surface anatomical landmark heat maps corresponding to these serial numbers. Figure 1 Finally, the processed body surface anatomical landmark heat map 2 is processed according to formula (2) to obtain the final detection coordinates.

[0051] Although the specific implementation of the present invention is described in detail in conjunction with the drawings, it should not be understood as limiting the scope of protection of this patent. Within the scope described in the claims, various modifications and variations that can be made by those skilled in the art without creative work are still within the scope of protection of this patent.

Claims

1. A two-stage method of photographic body position recognition and body surface anatomical landmark detection based on perspective assistance, characterized in that: The following steps are involved: Si: Regroup the 23 photography positions and generate viewpoint labels; S2: Based on the network model pre-trained with the ImageNet dataset as the initial weight, the photographic position classification phase network including the backbone network, the photographic position branch, the view classification branch, and the body surface anatomical landmark detection is trained; S3: Based on the network model pre-trained with the ImageNet dataset as the initial weight, the network for body surface anatomical landmark detection is trained; S4: Input the test data in the data set into the optimized two-stage network to obtain the photographic position classification results and the body surface anatomical landmark heat map; S5: Decode the two sets of heat maps into coordinates, use the average Gaussian similarity to evaluate the two sets of coordinates, and obtain a sequence of landmark types with high detection performance; S6: Using the high detection performance landmark point type sequence as prior information, performing selection processing on the two sets of heat maps obtained in step S4, and outputting the coordinate results; The structure of the view classification branch in step S2 is: first, N 41 , N 42 , N 43 , N 44 Feed them into the bottleneck layer of ResNet to increase their channels to 128, 256, 512, and 1024, respectively; then use a 3×3 convolution with 256 channels and a stride of 2 to downsample the high-resolution feature map, and then add it to the second-highest resolution feature map to fuse it into a low-level feature; After repeating this process three times, a minimum resolution feature map with 1024 channels is obtained. Then, the number of feature map channels is increased to 2048 through a 1×1 convolutional layer. Finally, the feature map is passed through an average pooling layer and a fully connected layer in turn to obtain the result of view 6 classification. The photographic posture classification branch in step S2 is the main branch of the photographic posture classification stage. It adopts a network structure similar to that of the viewpoint classification branch. The difference is that an SE module with a compression factor of 8 is added to recalibrate the feature maps of different resolutions. First, the SE module is inserted into the feature map with the highest resolution whose channels are increased to 128, and then the channel of the highest resolution feature map is adjusted. Secondly, the SE module is added to the secondary resolution feature map obtained by adding and fusion of the high-resolution feature map after downsampling and the secondary resolution feature map, and then recalibration is performed after the fusion of multiple resolution features; finally, the SE module is added to the lowest resolution feature map with 2048 channels obtained by 1×1 convolution, and the last channel feature recalibration before classification is performed; finally, the recalibrated feature map passes through the average pooling layer and the fully connected layer in turn to obtain the results of 23 classification of photographic postures.

2. The two-stage method of view-assisted photographic body position recognition and body surface anatomical landmark detection according to claim 1, characterized in that: In step S1, the 23 photographic postures are analyzed and regrouped according to the viewing angle of the human body in the image, that is, the relative angle between the human body and the camera.

3. The two-stage method of view-assisted photographic body position recognition and body surface anatomical landmark detection according to claim 1, characterized in that: In step S2, the height or width of the human body bounding box is expanded to a fixed aspect ratio of height: width = 4:3, the bounding box is cropped from the image, and then it is resized to a fixed size of 384 pixels × 288 pixels as the input image. The image is input to the network in the photography posture stage. The backbone network uses HRNetV2-W48 to extract feature maps N of four resolutions. 41 ∈R N*ω*H / 4*W / 4 , N 42 ∈R N*2ω*H / 8*W / 8 , N 43 ∈R N*4ω*H / 16*W / 16 , N 44 ∈R N*8ω*H / 32*W / 32 , w is set to 48, H and W are the height and width of the input image.

4. The two-stage method of view-assisted photographic body position recognition and body surface anatomical landmark detection according to claim 1, characterized in that: The loss of the view classification branch in step S2 is L poscls , using cross entropy loss; the loss of the photography position branch is L angcls , using cross entropy loss; the loss of the anatomical landmark detection branch is L landmark , using mean square error loss; the loss in the training phase is set to: L=0.00001*L poscls +0.00001*L angcls +L landmark (1) Random rotation and random scaling are used for image data enhancement. The random rotation range is ±45 degrees, and the random scaling range is ±35%. Based on the idea of ​​transfer learning, the network model pre-trained on ImageNet is used for model migration on the training set in the dataset. The Adam optimizer is used for training optimization. The basic learning rate is set to 1e-3, and it is reduced to 1e-4 and 1e-5 in the 70th and 100th epochs respectively. The training process ends within 110 epochs.

5. The two-stage method of view-assisted photographic body position recognition and body surface anatomical landmark detection according to claim 1, characterized in that: The stage of detecting anatomical landmarks on the body surface in step S3 includes the following steps: The height or width of the human body bounding box is expanded to a fixed aspect ratio of height: width = 4:

3. The bounding box is cropped from the image and then resized to a fixed size of 384 pixels × 288 pixels as the input image. The image is input to the network in the photography posture stage. First, HRNetV2-W48 is used as the backbone network to extract feature maps N of four resolutions. 41 , N 42 , N 43 , N 44 ; Then N 41 Predict the 80-channel heat map through 1×1 convolution, using L landmark As the loss function of this stage, the data augmentation method, pre-training model and optimizer selection in the training stage are the same as those in the photographic position classification stage.

6. The two-stage method of view-assisted photographic body position recognition and body surface anatomical landmark detection according to claim 1, characterized in that: The body surface anatomical landmark heat map in step S4 includes the body surface anatomical landmark heat map 1 output by the network in the photographic posture classification stage in step S2 and the body surface anatomical landmark heat map 2 output by the network in the body surface anatomical landmark detection stage in step S3.

7. The two-stage method of view-assisted photographic body position recognition and body surface anatomical landmark detection according to claim 1, characterized in that: The two sets of coordinates are evaluated in step S5 using the average Gaussian similarity, which is expressed as follows: Where i is the index of the landmark type; d j is the predicted coordinates of the jth landmark of the i-th type of anatomical landmark, each landmark will appear only once in each human instance, and the Euclidean distance between the corresponding true value coordinates; s is the scale factor, which is the size of the bounding box of the human instance where the i-th key point of type i is located; κ i is a constant that controls the attenuation rate of the i-th type of anatomical surface landmark; N is the total number of i-type landmark points in the test image.

8. The two-stage method of view-assisted photographic body position recognition and body surface anatomical landmark detection as claimed in claim 1, characterized in that: The selection processing and coordinate extraction in step S6 includes the following steps: using the photographic position classification results of the photographic position classification stage to filter the surface anatomical landmark heat map 2; if the serial numbers corresponding to the channels that are not 0 in the filtered surface anatomical landmark heat map 2 contain a high detection performance landmark point type sequence, then the channels in the surface anatomical landmark heat map 2 corresponding to these serial numbers are replaced with the channels in the surface anatomical landmark heat map 1 corresponding to these serial numbers; finally, the surface anatomical landmark heat map 2 after the aforementioned processing is processed according to the following formula to obtain the final detection coordinates.

Images