Student behavior detection method based on classroom information guiding anchor frame

Abstract

The invention discloses a student behavior detection method based on a classroom information guide anchor frame. The method comprises the steps of collecting data; constructing a student behavior detection network GASP-RPN based on the classroom information guide anchor frame; carrying out GASP-RPN training and parameter learning on a student behavior detection network; and inputting the images of the verification set into the trained student behavior detection network GASP-RPN, and identifying and classifying the actions of the students. According to the invention, the GASP-RPN uses the structure information of the classroom scene, so that the behaviors of the students can be detected more accurately. The GASP-RPN generates anchor frames of different scales and shapes through the learnable anchor frame network, the multi-scale detection problem is effectively solved, the detection capability of targets of different sizes is improved, the adaptability of the model to the target shape and position change is improved through the feature adaptive module and the deformable convolutional network, and thus the robustness of the detection system is enhanced.

Description

Technical Field

[0001] The present invention belongs to the technical field of target detection, and in particular relates to a student behavior detection method based on classroom information-guided anchor frames. Background Art

[0002] The classroom scene environment is complex, and there are many factors that affect the detection of classroom student behavior. Among the many influencing factors, the multi-scale problem has a particularly significant impact. Current research mainly applies the target algorithm of general scenes directly to the classroom scene to solve the detection difficulties, ignoring the particularity of the classroom scene and failing to solve the problem well.

[0003] Object detection methods can be roughly divided into two-stage object detection methods and single object detection methods. The core of these methods is to generate high-quality anchor boxes. Currently, there are three main ways to generate anchor boxes:

[0004] (1) Sliding window anchors. By generating anchors in the feature map in a sliding window manner, the size and shape of the target object can be captured, improving the accuracy and generalization ability of the target detection model. This method of generating anchors is basically used in anchor-based detectors, that is, candidate regions are provided in this way in the first stage. For example, Faster RCNN uses the candidate region proposal network RPN (Region Proposal Networks) to generate candidate regions. It uses a small fully convolutional network to map each sliding window to the original image corresponding to the feature map, and generates anchor frames of 3 scales and 9 ratios at the corresponding position of the image to frame the real target. Then, the non-maximum suppression strategy NMS (Non-Maxium-Suppression) is combined with the intersection-over-union threshold to remove some anchor frames, and the anchor frame with the highest score is retained as the candidate frame;

[0005] (2) The cascade architecture generates anchors. First, the cascade architecture generates anchors at the bottom layer of the network and proposes a series of candidate regions. These candidate regions are then sent to the subsequent network stage for target classification and bounding box regression. Usually, the cascade architecture uses anchor boxes of different scales and proportions at multiple levels or stages to cover targets of different sizes and shapes. Existing studies have attempted to use the cascade architecture to reject samples that are prone to appearing in the early stages, and then iteratively regress the bounding boxes for gradual refinement. Specifically, this method gradually improves the detection performance through multiple cascade models or stages. In the early stages, the model may reject some samples that are prone to misdetection in order to reduce erroneous detections; in the subsequent stages, the model will iteratively regress and refine the bounding boxes to help the model better distinguish between targets and backgrounds, improve the accuracy of target positioning and classification, effectively reduce positioning errors, and improve the robustness and accuracy of target detection;

[0006] (3) Single-stage object detection algorithms Compared with two-stage methods, single-stage object detection omits the proposal step but still uses a sliding window to generate anchor boxes. For example, the object detection algorithms SSD and DenseBox generate dense anchors in the feature map and evaluate them like a multi-class RPN. The YOLOv2 algorithm uses sliding window anchors for classification and spatial position prediction to obtain a higher recall rate than before. These single-stage methods use densely generated anchors to cover objects of different scales and proportions, and achieve object detection by evaluating these anchors.

[0007] The above anchor point-based anchor box detection method has achieved good detection accuracy in many scenarios, but the detection performance of this method mainly depends on providing high-quality candidate boxes. The above sliding window to generate uniform anchor point anchor box solution is not the best method either. Uniform anchor points generate anchor boxes of predefined scale and shape, so that the scale and shape of the anchor box are fixed in advance, which may cause two problems:

[0008] ① A set of anchor boxes with fixed scales and shapes must be predefined for different problems. This predefined design may affect the speed and accuracy of the detection model;

[0009] ② In order to maintain a sufficiently high recall rate, a large number of neat anchors are provided, but most of the candidate boxes are false candidate boxes that are irrelevant to the real target. At the same time, a large number of anchors will also lead to significant computational costs, especially in the proposal stage, which involves a large number of classifiers for foreground and background discrimination. Summary of the invention

[0010] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a student behavior detection method based on a classroom information guided anchor frame with high accuracy, strong robustness and improved computing resource efficiency.

[0011] The technical solution adopted to solve the above technical problems is: a student behavior detection method based on classroom information guidance anchor frame, including the following steps:

[0012] Step 1. Data Collection

[0013] Collect student behavior pictures in real classrooms and annotate the behaviors to form a dataset, and divide the dataset into a training set and a validation set;

[0014] Step 2. Construct the student behavior detection network GASP-RPN based on classroom information-guided anchor frame

[0015] The student behavior detection network GASP-RPN is composed of a feature extraction module, a classroom structure point guide anchor frame generation module, and a positioning target category and position module connected in sequence;

[0016] The feature extraction module is used to extract features from the input image through a deep neural network;

[0017] The classroom structure point guided anchor frame generation module is used to generate multi-scale anchor frames and feature maps;

[0018] The target category and position positioning module is used to align the anchor frame with the feature map and regress the target position and shape respectively after passing through the fully connected layer, and output the category through the Softmax function;

[0019] The classroom structure point guided anchor frame generation module is composed of a parallel classroom structure discovery module and a learnable anchor frame module; the classroom structure discovery module is composed of a deep neural network and a clustering module, the deep neural network is used to obtain the structural points constituting the row and column layout in the classroom from the input feature map, and the clustering module is used to group the structural points and fit the classroom row and column lines;

[0020] The learnable anchor frame module is composed of an anchor frame generation module and a feature adaptation module. The anchor frame generation module has two branches. Branch one is composed of a 1×1 convolution layer, and the channel mapping of the volume layer is W×H×1, where W is the width of the feature map and H is the height of the feature map, and is used to generate a position prediction probability distribution map. Branch two is composed of a 1×1 convolution layer, and the channel mapping of the volume layer is W×H×2, and is used to achieve consistency and alignment of the anchor through the anchor shape prediction branch. The feature adaptation module is composed of an offset module, and is used to adapt to features based on changing anchor geometry.

[0021] Step 3. Student behavior detection network GASP-RPN training and parameter learning

[0022] Step 3.1. Initialize network model parameters: Use the ResNet101 pre-trained model on the ImageNet dataset to initialize the parameters of the feature extraction module;

[0023] Step 3.2. Use the training set to train the student behavior detection network GASP-RPN, and use the loss function L as the objective function to train and optimize the learning parameters and save the network model;

[0024] Step 3.3. Use the stochastic gradient descent algorithm to optimize the objective function, and use the back propagation algorithm to update the parameters of the student behavior detection network GASP-RPN until the value of the objective function no longer decreases, and then end the training and save the trained student behavior detection network GASP-RPN;

[0025] Step 4. Input the images of the validation set into the trained student behavior detection network GASP-RPN to identify and classify the student's actions;

[0026] Step 4.1. The feature extraction module extracts features from the input image and outputs a feature map f q ;

[0027] Step 4.2. The classroom structure point guides the anchor frame generation module to input the feature map f q Processing, output anchor boxes and adaptive feature maps f′q of different scales;

[0028] Step 4.3. The module for locating the target category and position uses RoI Align to align the anchor box and the adaptive feature map f′q. After passing through the fully connected layer, it outputs the target category, target position and shape (x′, y′, w, h), where (x′, y′) is the target center coordinate, w is the target width, and h is the target height, and the classification result of the student's action is obtained.

[0029] As a preferred technical solution, the behaviors in step 1 include raising the head, lowering the head, turning the head, playing with a mobile phone, reading a book, raising the hand, and lying down.

[0030] As a preferred technical solution, the feature extraction module uses ResNet-50 / 101 as the backbone network.

[0031] As a preferred technical solution, the loss function L is composed of the classification loss L cls , regression loss L reg , anchor point positioning loss L loc , anchor box scale and shape loss L shape According to the following formula,

[0032] L=λ 1 L loc +λ 2 L shape +L cls +L reg

[0033] In the formula, λ 1 ,λ 2 Denotes the anchor point positioning loss L loc and anchor box scale and shape prediction loss L shape The scaling factor.

[0034] As a preferred technical solution, the classroom structure point guide anchor frame generation module in step 4.2 generates the input feature map f q The processing method is:

[0035] 1) The classroom structure discovery module inputs the feature map f qThrough YOLO algorithm analysis, the structural point set that constitutes the row and column layout of the classroom is obtained. The structural point set includes the features of the student's head and the edge of the desk. At the same time, the 1-nearest neighbor algorithm is used to group the structural points in the structural point set, fit the classroom row and column lines, and obtain the classroom structural point coordinates (x i ,y i ), i = 0, 1, ..., k, k is a positive integer, determine the location of students in the classroom, and find the nearest neighbor structure point (x) of each classroom structure point according to the nearest neighbor distance algorithm. j ,y j ), the nearest neighbor distance d can be obtained by the following formula min ;

[0036]

[0037] 2) Through the classroom structure point and the nearest neighbor distance d min Determine the coordinates of the upper left corner of the target area (x a ,y a ) and the coordinates of the lower right corner (x d ,y d ),

[0038]

[0039]

[0040] Where α and β are the scale transformation factors of the target area;

[0041] 3) In the target area, the learnable anchor frame module generates anchor frames of multiple sizes according to the classroom structure points;

[0042] In the target area, the position and shape of the target are represented by a 4-tuple in the form of (x′, y′, w, h), where (x′, y′) is the target center coordinate, w is the target width, and h is the target height. The probability distribution of the target position and shape p(x′, y′, w, h|r) is:

[0043] p(x′,y′,w,h|r)=p(x′,y′|r)·p(w,h|x′,y′,r)

[0044] Where p(x′,y′|r) represents the probability of predicting the center position of the target, i.e. the anchor point, in the target region, and p(w,h|x′,y′,r) represents the probability of predicting the scale and shape of the anchor box given the center position of the target in the target region.

[0045] 4) The anchor-guided feature adaptation module predicts an offset, applies the deformable convolution and offset to the original feature map, and obtains the adaptive feature map f′ through the following formula:q ,

[0046] f′ q =N T (f q ,w q ,h q )

[0047] Where N T (·) is a 3×3 deformable convolution, f q is the feature at the qth position, (w q ,h q ) is the shape of the anchor box corresponding to the qth position.

[0048] The beneficial effects of the present invention are as follows:

[0049] The GASP-RPN of the present invention utilizes the structural information of the classroom scene to detect student behavior more accurately. GASP-RPN generates anchor frames of different scales and shapes through a learnable anchor frame network, effectively solving the multi-scale detection problem and improving the detection capability of targets of different sizes. The adaptability of the model to target shape and position changes is improved through the feature adaptation module and deformable convolutional network, thereby enhancing the robustness of the detection system.

[0050] Compared with traditional target detection methods, GASP-RPN reduces unnecessary calculations and improves the efficiency of computing resources by generating more accurate anchor boxes. By generating high-quality candidate boxes, the recall rate of target detection is improved, ensuring that more real targets are detected. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 These are 7 representative student classroom behavior diagrams of the present invention.

[0052] Figure 2 It is a structural diagram of the student behavior detection network GASP-RPN of the present invention.

[0053] Figure 3 It is a schematic diagram of the target existence area of ​​the present invention.

[0054] Figure 4 Schematic diagram of the learnable anchor frame network of the present invention.

[0055] Figure 5 This is a comparison diagram of the IOU of candidate boxes generated by the method of the present invention and the existing method. DETAILED DESCRIPTION

[0056] The present invention will be further described in detail below in conjunction with the accompanying drawings and examples, but the present invention is not limited to the following embodiments.

[0057] The student behavior detection method based on classroom information guidance anchor frame of this embodiment includes the following steps:

[0058] Step 1. Data Collection

[0059] We collected about 2,000 minutes of college classroom videos authorized by a university in Henan Province. All the video contents were spontaneous behaviors of students in real classrooms. We summarized 7 representative student classroom behaviors, such as Figure 1 As shown, including looking up, lowering, turning the head, playing with mobile phones, reading, raising hands, and lying down, the data was then preprocessed and video frames were extracted at 35-second intervals to form 1920*1080 pixel JPG format images. Subsequently, the Labelme tool was used to label these images with 7 types of student classroom behavior categories. All labeled data was stored in JSON format, including image information, annotation box coordinates, and corresponding labels. The images and corresponding labeled data were divided into training sets and validation sets for distribution, see Table 1;

[0060] Table 1. Training set and validation set

[0061]

[0062] Step 2. Construct the student behavior detection network GASP-RPN based on classroom information-guided anchor frame

[0063] The student behavior detection network GASP-RPN is composed of a feature extraction module, a classroom structure point guide anchor frame generation module, and a positioning target category and position module. Figure 2 ;

[0064] The feature extraction module uses ResNet-50 / 101 as the backbone network to extract features from the input image through a deep neural network;

[0065] The classroom structure point guided anchor frame generation module is used to generate multi-scale anchor frames and feature maps;

[0066] The target category and position positioning module is used to align the anchor frame with the feature map and regress the target position and shape after passing through the fully connected layer, and output the category through the Softmax function;

[0067] The classroom structure point guided anchor frame generation module is composed of a parallel classroom structure discovery module and a learnable anchor frame module; the classroom structure discovery module is composed of a deep neural network and a clustering module. The deep neural network is used to obtain the structural points that constitute the row and column layout in the classroom from the input feature map, and the clustering module is used to group the structural points and fit the classroom row and column lines;

[0068] like Figure 3The learnable anchor frame module consists of an anchor frame generation module and a feature adaptation module. The anchor frame generation module has two branches, branch one N S It consists of a 1×1 convolutional layer. The channel mapping of the volume base is W×H×1, where W is the width of the feature map and H is the height of the feature map. It is used to generate the position prediction probability distribution map. T It consists of a 1×1 convolutional layer whose channel mapping is W×H×2, which is used to achieve consistency and alignment of anchors through anchor shape prediction branches. The feature adaptation module consists of an offset module for adapting to features based on changing anchor geometry.

[0069] Step 3. Student behavior detection network GASP-RPN training and parameter learning

[0070] Step 3.1. Initialize network model parameters: Use the ResNet101 pre-trained model on the ImageNet dataset to initialize the parameters of the feature extraction module;

[0071] Step 3.2. Use the training set to train the student behavior detection network GASP-RPN, use the loss function L as the objective function, train and optimize the learning parameters and save the network model;

[0072] Among them, the loss function L is composed of the classification loss L cls , regression loss L reg , anchor point positioning loss L loc , anchor box scale and shape loss L shape According to the following formula,

[0073] L=λ 1 L loc +λ 2 L shape +L cls +L reg

[0074] Where λ 1 ,λ 2 Denotes the anchor point positioning loss L loc and anchor box scale and shape prediction loss L shape The scaling factor, λ 1 =0.1,λ 2 =0.2.

[0075] Step 3.3. Use the stochastic gradient descent algorithm to optimize the objective function, and use the back propagation algorithm to update the parameters of the student behavior detection network GASP-RPN until the value of the objective function no longer decreases, and then end the training and save the trained student behavior detection network GASP-RPN;

[0076] Step 4. Input the images of the validation set into the trained student behavior detection network GASP-RPN to identify and classify the student's actions;

[0077] Step 4.1. The feature extraction module extracts features from the input image and outputs a feature map f q ;

[0078] Step 4.2. The classroom structure point guides the anchor frame generation module to input the feature map f q Processing, output anchor boxes and adaptive feature maps f′ of different scales q , the specific operations are as follows:

[0079] 1) The classroom structure discovery module inputs the feature map f q Through YOLO algorithm analysis, the structural point set that constitutes the row and column layout of the classroom is obtained. The structural point set includes the features of the student's head and the edge of the desk. At the same time, the 1-nearest neighbor algorithm is used to group the structural points in the structural point set, fit the classroom row and column lines, and obtain the classroom structural point coordinates (x i ,y i ), i = 0, 1, ..., k, k is a positive integer, determine the location of students in the classroom. According to the nearest neighbor distance algorithm, find the nearest neighboring structure point (x j ,y j ), the nearest neighbor distance d can be obtained by the following formula min ;

[0080]

[0081] 2) Through the classroom structure point and the nearest neighbor distance d min Determine the coordinates of the upper left corner of the target area (x a ,y a ) and the coordinates of the lower right corner (x d ,y d ),

[0082]

[0083] In the formula, α and β are the scale transformation factors of the target area, α = 2, β = 1.5; based on the classroom structure information obtained, any classroom picture can be input to determine the target area and non-target area in the classroom picture. Figure 4 , the area marked by the red box is the target area r, and the area outside the box is the non-target area.

[0084] 3) In the target region r, the anchor frame module can be learned to generate anchor frames of multiple sizes according to the classroom structure points

[0085] In the target region r, the position and shape of the target are represented by a 4-tuple in the form of (x′, y′, w, h), where (x′, y′) is the target center coordinate, w is the target width, and h is the target height. The probability distribution of the target position and shape p(x′, y′, w, h|r) is:

[0086] p(x′,y′,w,h|r)=p(x′,y′|r)·p(w,h|x′,y′,r)

[0087] Where p(x′,y′|r) represents the probability of predicting the center position of the target, i.e. the anchor point, in the target region r, and p(w,h|x′,y′,r) represents the probability of predicting the scale and shape of the anchor box given the center position of the target in the target region r.

[0088] 4) The anchor-guided feature adaptation module predicts an offset, applies the deformable convolution and offset to the original feature map, and obtains the adaptive feature map f′ through the following formula: q ,

[0089] f′ q =N T (f q ,w q , h q)

[0090] Where N T (·) is a 3×3 deformable convolution, f q is the feature at the qth position, (w q ,h q ) is the corresponding anchor box shape;

[0091] Step 4.3. The target category and position module locates the anchor box and the adaptive feature map f′ q Use RoI Align to align features. After passing through the fully connected layer, the target category, target position and shape (x′, y′, w, h) are output, where (x′, y′) is the target center coordinate, w is the target width, and h is the target height, and the classification result of the student's action is obtained.

[0092] experiment

[0093] In order to verify the beneficial effects of the present invention, the inventors conducted comparative experiments using the method of Example 1 and the existing methods FasterRCNN, YOLOv5, YOLO-MS, Faster RCNN+DPS, and Faster RCNN+CABM algorithms. The detection accuracy of various models is shown in Table 2.

[0094] Table 2. Detection accuracy of various models

[0095]

[0096] GASP-RPN 88.3% 89.2% 80.1% 71.3% 88.6% 94.5% 87.4% 85.6%

[0097] Experimental results: Figure 5 In the comparison of the generated candidate boxes IOU, it is found that the learnable anchor box generation network LABN can generate higher quality candidate boxes than the traditional RPN network, which is mainly manifested in two aspects: ① more candidate boxes; ② a larger proportion of high IOU candidate boxes. This method improves the accuracy of target detection by increasing the number of anchor boxes compared with other comparative target detection algorithms, as shown in Table 2. The detection accuracy of GASP-RPN is higher than that of other detection methods, fully considering the classroom scene information, effectively solving the multi-scale problem, and is more suitable for the detection of students' classroom behavior in classroom scenes.

Claims

1. A student behavior detection method based on classroom information guided anchor frame, characterized in that: The following steps are involved: Step 1. Data Collection Collect student behavior pictures in real classrooms and annotate the behaviors to form a dataset, and divide the dataset into a training set and a validation set; Step 2. Construct the student behavior detection network GASP-RPN based on classroom information-guided anchor frame The student behavior detection network GASP-RPN is composed of a feature extraction module, a classroom structure point guide anchor frame generation module, and a positioning target category and position module connected in sequence; The feature extraction module is used to extract features from the input image through a deep neural network; The classroom structure point guided anchor frame generation module is used to generate multi-scale anchor frames and feature maps; The target category and position positioning module is used to align the anchor frame with the feature map and regress the target position and shape respectively after passing through the fully connected layer, and output the category through the Softmax function; The classroom structure point guided anchor frame generation module is composed of a parallel classroom structure discovery module and a learnable anchor frame module; the classroom structure discovery module is composed of a deep neural network and a clustering module, the deep neural network is used to obtain the structural points constituting the row and column layout in the classroom from the input feature map, and the clustering module is used to group the structural points and fit the classroom row and column lines; The learnable anchor frame module is composed of an anchor frame generation module and a feature adaptation module. The anchor frame generation module has two branches. Branch one is composed of a 1×1 convolution layer, and the channel mapping of the volume layer is W×H×1, where W is the width of the feature map and H is the height of the feature map, and is used to generate a position prediction probability distribution map. Branch two is composed of a 1×1 convolution layer, and the channel mapping of the volume layer is W×H×2, and is used to achieve consistency and alignment of the anchor through the anchor shape prediction branch. The feature adaptation module is composed of an offset module, and is used to adapt to features based on changing anchor geometry. Step 3. Student behavior detection network GASP-RPN training and parameter learning Step 3.

1. Initialize network model parameters: Use the ResNet101 pre-trained model on the ImageNet dataset to initialize the parameters of the feature extraction module; Step 3.

2. Use the training set to train the student behavior detection network GASP-RPN, and use the loss function L as the objective function to train and optimize the learning parameters and save the network model; Step 3.

3. Use the stochastic gradient descent algorithm to optimize the objective function, and use the back propagation algorithm to update the parameters of the student behavior detection network GASP-RPN until the value of the objective function no longer decreases, and then end the training and save the trained student behavior detection network GASP-RPN; Step 4. Input the images of the validation set into the trained student behavior detection network GASP-RPN to identify and classify the student's actions; Step 4.

1. The feature extraction module extracts features from the input image and outputs a feature map f q ; Step 4.

2. The classroom structure point guides the anchor frame generation module to input the feature map f q Processing, output anchor boxes and adaptive feature maps f′ of different scales q ; Step 4.

3. The target category and position module locates the anchor box and the adaptive feature map f′ q Use RoI Align to align features and output the target category, target position and shape (x ′ ,y ′ ,w,h), where (x ′ ,y ′ ) is the target center coordinate, w is the target width, h is the target height, and the classification result of the student’s action is obtained.

2. According to claim 1, the student behavior detection method based on classroom information guidance anchor frame is characterized in that: The behaviors described in step 1 include raising the head, lowering the head, turning the head, playing with the mobile phone, reading a book, raising the hand, and lying down.

3. The student behavior detection method based on classroom information guidance anchor frame according to claim 1 is characterized in that: The feature extraction module uses ResNet-50 / 101 as the backbone network.

4. The student behavior detection method based on classroom information guidance anchor frame according to claim 1 is characterized in that: The loss function L is composed of the classification loss L cls , regression loss L reg , anchor point positioning loss L loc , anchor box scale and shape loss L shape According to the following formula, L=λ1L loc +λ2L shape +L cls +L reg Where λ1, λ2 represent the anchor point positioning loss L loc and anchor box scale and shape prediction loss L shape The scaling factor.

5. The student behavior detection method based on classroom information guidance anchor frame according to claim 1 is characterized in that: The classroom structure point guide anchor frame generation module described in step 4.2 is used to generate the input feature map f q The processing method is: 1) The classroom structure discovery module inputs the feature map f q Through YOLO algorithm analysis, the structural point set that constitutes the row and column layout of the classroom is obtained. The structural point set includes the features of the student's head and the edge of the desk. At the same time, the 1-nearest neighbor algorithm is used to group the structural points in the structural point set, fit the classroom row and column lines, and obtain the classroom structural point coordinates (x i ,y i ), i = 0, 1, ..., k, k is a positive integer, determine the location of students in the classroom, and find the nearest neighbor structure point (x) of each classroom structure point according to the nearest neighbor distance algorithm. j ,y j ), the nearest neighbor distance d can be obtained by the following formula min ; 2) Through the classroom structure point and the nearest neighbor distance d min Determine the coordinates of the upper left corner of the target area (x a ,y a ) and the coordinates of the lower right corner (x d ,y d ), Where α and β are the scale transformation factors of the target area; 3) In the target area, the learnable anchor frame module generates anchor frames of multiple sizes according to the classroom structure points; In the target area, the position and shape of the target are represented by a 4-tuple in the form of (x′, y′, w, h), where (x′, y′) is the target center coordinate, w is the target width, and h is the target height. The probability distribution of the target position and shape p(x′, y′, w, h|r) is: p(x′,y′,w,h|r)=p(x′,y′|r)·p(w,h|x′,y′,r) Where p(x′, y′|r) represents the probability of predicting the center position of the target, i.e. the anchor point, in the target existence area, and p(w, h|x′, y′, r) represents the probability of predicting the scale and shape of the anchor box given the center position of the target in the target existence area. 4) The anchor-guided feature adaptation module predicts an offset, applies the deformable convolution and offset to the original feature map, and obtains the adaptive feature map f′ through the following formula: q , f′ q =N T (f q ,w q ,h q ) Where N T (·) is a 3×3 deformable convolution, f q is the feature at the qth position, (w q ,h q ) is the shape of the anchor box corresponding to the qth position.

Images