A camouflaged target frame detection method based on ternary cascade perception

Abstract

The present invention discloses a camouflaged target frame detection method based on ternary cascade perception, and constructs a cascade perception network in combination with the characteristics of the camouflaged target. The input to the output respectively includes a feature extraction module, a ternary cascade perception module and a detection output head. The feature extraction module obtains multi-scale features by cascading a feature extractor and a feature pyramid network; the ternary cascade perception module includes a feature perception module, a key point perception module and a region perception module; the detection output head performs back propagation by updating a joint loss function during training, and outputs the feature map encoding as a target category, target frame information and target confidence during detection. The cascade perception network designed by the present invention further improves the ability to accurately detect and locate camouflaged targets by learning the relationship information between features, key point location information and multi-scale information of key areas, and has high detection accuracy and reliability.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer vision and relates to a camouflage target frame detection method based on ternary cascade perception. Background Art

[0002] In a narrow sense, camouflaged object detection aims to find targets camouflaged in a given image, such as chameleons hiding in the environment by changing the color of their skin, or armed soldiers lurking in the mountains. In a broad sense, camouflaged object detection aims to detect any low-detectability objects in an image, such as component defect detection in industry, or lesion sign detection in pathology. The research on camouflaged object detection is not only an important scientific issue in the field of visual perception, but also can bring huge economic and social benefits.

[0003] However, the color and pattern of camouflaged objects are almost integrated with the environment, which greatly reduces the probability of being detected, identified or locked. Currently, researchers have conducted corresponding research on camouflaged target recognition.

[0004] Some scholars use other technologies to identify visible light camouflaged targets. For example, Zhang Liguo and others from Yanshan University combined hyperspectral technology to identify camouflaged targets and applied for a patent titled "Method for realizing hyperspectral camouflaged target identification based on improved OIF and SVM". Gu Guohua and others from Nanjing University of Science and Technology used polarization technology and applied for a patent titled "Method for identifying camouflaged targets based on K-means clustering and polarization information extraction". However, the combination of multiple technologies has failed to directly solve the problem of camouflaged target identification in the visible light band.

[0005] In the field of camouflaged target recognition in the visible light band, in 2020, Zhou Jing and others from Jinan University conducted research on the continuous frame segmentation of camouflaged targets and published a paper titled "Camouflaged Moving Target Detection Based on Optical Flow Field Segmentation". Deng-Ping Fan and others from the UAE Institute of Artificial Intelligence published a paper titled "Camouflaged Object Detection" at the IEEE Conference on Computer Vision and Pattern Recognition, proposing an effective camouflaged target segmentation algorithm SINet. In 2021, Jinnan Yan from the University of Dayton, USA, inspired by biology, proposed a new simple and efficient network framework MirrorNet, which greatly improved the existing camouflaged object recognition performance. Yunqiu Lv from Northwestern Polytechnical University proposed an integrated network that can locate, segment and sort camouflaged targets. Cai Wei and others from the Rocket Force Engineering University of the Chinese People's Liberation Army applied for a patent titled "A Visible Light Camouflaged Target Recognition Method Based on Magnifying Glass Observation Effect", which simulates the magnifying glass observation effect to obtain high-quality visible light camouflaged target segmentation maps. The above methods are all committed to improving the camouflaged target segmentation technology, so the network structure is generally complex and time-consuming.

[0006] Compared with the camouflaged target segmentation technology mentioned above, the camouflaged target frame detection technology is faster and more practical. It only uses the smallest box that completely contains the camouflaged target to mark the camouflaged target, and there is no need to perform pixel-by-pixel classification, so the detection time is greatly reduced. At present, there are not many studies in the field of camouflaged target frame detection. Cao Tieyong and others from the Army Engineering University have realized target frame detection by improving the mature YOLO algorithm and applied for a patent entitled "A Camouflaged Target Detection Method Based on Improved YOLO Algorithm". However, the YOLO algorithm that this method relies on is a general target detection algorithm, which has poor targeting for camouflaged targets.

[0007] In summary, researchers currently usually study camouflage target segmentation technology, and most of them use other technologies to do so. There are not many studies on camouflage target frame detection in the visible light band. Therefore, in order to make up for the shortcomings of the existing camouflage target recognition field, an accurate and fast camouflage target frame detection method must be proposed. Summary of the invention

[0008] In order to make up for the shortcomings of the existing frame detection disguised target technology and achieve accurate implementation of disguised target frame detection, the present invention provides a disguised target frame detection method based on ternary cascade perception.

[0009] The inventive concept and specific steps of the present invention are described as follows:

[0010] The basic concept of the present invention is: in order to accurately detect the camouflaged target in the image and solve the problem that the key information of the camouflaged target is difficult to obtain, the network structure is designed, and the cascade perception network CPNet is specifically constructed. The algorithm innovation lies in the proposal of the ternary cascade perception module (TCPM), which focuses on extracting the relationship information between features, the location information of key points, and the multi-scale information of key areas. At the same time, a joint loss function is proposed, and different loss functions are used for the prediction of target category, target frame, and target confidence, and the training loss value is obtained jointly.

[0011] The present invention discloses a camouflaged target frame detection method based on ternary cascade perception, comprising the following steps:

[0012] Step 1: Construct a ternary cascade perception module (TCPM) based on the characteristics of the camouflaged target.

[0013] TCPM is built based on the characteristics of camouflaged targets, including the feature perception module (FPM), the key point perception module (KPM) and the region perception module (RPM). The three modules are embedded in the backbone network in a cascade manner.

[0014] The feature perception module FPM obtains the relative relationship information between features through the graph convolution network (GCN). For the input tensor F∈R W×HxC First, use the convolution layer combination CBR (convolution Conv + batch normalization BN + Relu activation function) for preprocessing to obtain the feature map

[0015]

[0016] Among them, Conv(BN(Relu(·))) represents the convolution layer combined CBR operation, and then the whole feature map is pooled in two dimensions through global average pooling (GAP) to generate a three-dimensional tensor pooling feature map F GAP ∈R 1×1×C :

[0017]

[0018] Among them, GAP(·, (W=1, H=1)) represents the operation of globally pooling the input feature map into a feature map with W=1 and H=1. In the convolutional neural network, the number of channels C represents the number of features extracted from the feature map, so the three-dimensional tensor F GAP ∈R 1×1×C It can be regarded as a set of 256 feature points. In order to further obtain the relative relationship information between features, the graph convolutional network GCN is used to learn the relationship between nodes and obtain a new three-dimensional tensor feature map F with node relationships. GCN ∈R 1×1×C Based on this, a Softmax normalization operation is performed on the channel dimension to normalize the relationship information to 0-1, and a three-dimensional tensor weight feature map is obtained.

[0019]

[0020] Among them, Softmax(·, dim=C) represents the Softmax operation on the channel dimension, and the weight feature map F Weight With preprocessed feature map Multiply channel by channel to obtain the relation-weighted feature map Finally, a residual connection is performed with the original input, and pixel-by-pixel addition is performed to obtain the output feature tensor of the feature perception module FPM Re-enter the network backbone.

[0021]

[0022] The key point perception module obtains the location information of the key points through average pooling (AP) of different dimensions. For the output tensor of the feature perception module First, use the convolutional layer combined with CBR for preprocessing to obtain the feature map

[0023]

[0024] Unlike the feature perception module FPM, average pooling does not operate on the entire feature map, but divides it into two dimensions, W and H, to obtain the horizontal and vertical position information of the feature nodes in the feature map, and generate a pair of three-dimensional tensor pooling feature maps F Pool-H ∈R W×1×C and F Pool-W ∈R 1×H×C :

[0025]

[0026]

[0027] Secondly, the two feature maps containing the position information of the key points in different directions are multiplied to obtain the feature map F with the global position information of the key points embedded. W×H ∈R W×H×C Then, the convolutional layer is combined with CBR to fuse the information between channels, and then the position information is normalized to 0-1 through the Softmax normalization operation of the channel dimension to obtain the weight feature map of the key point position information of the three-dimensional tensor

[0028]

[0029] The weight feature map With preprocessed feature map Multiply channel by channel to obtain the relation-weighted feature map Finally, the feature tensor output by the input (i.e., the feature perception module FPM) ) is connected with residuals, and the output feature tensor of the key point perception module KPM is obtained by adding pixel by pixel. Re-enter the network backbone.

[0030]

[0031] The region perception module increases the receptive field by fusing dilated convolutions (DConv) of different scales to obtain more multi-scale information of key regions. For the output tensor of the key point perception module First, use the convolutional layer combined with CBR for preprocessing to obtain the feature map

[0032]

[0033] Different from the key point perception module KPM, in order to obtain multi-scale information of key areas, it is necessary to perform feature extraction in a small range. Two dilated convolution branches of different sizes are used to expand the receptive field and obtain multi-scale regional information. The dilated convolution kernel sizes are 3×3 and 5×5 respectively, and the dilated coefficients are both 2. Then the two sets of feature maps are superimposed to obtain the receptive field enhanced feature map F with twice the number of channels. RF ∈R W×H×2C :

[0034]

[0035] Among them, DConv(·, K=i, D=2) represents the dilated convolution with a dilated coefficient of 2 and a convolution kernel size of i. Add(·, ·) represents the superposition operation of the two sets of feature maps in the channel dimension, and then the Softmax operation is performed on the channel dimension. Then, the convolution layer is combined with CBR to perform channel recovery and information fusion. Then, the position information is normalized to 0~1 through the Softmax normalization operation in the channel dimension, and the multi-scale information weight feature map of the three-dimensional tensor region is obtained.

[0036]

[0037] The weight feature map With preprocessed feature map Multiply channel by channel to obtain the relation-weighted feature map Finally, with the module input (i.e., the output feature tensor of the key point perception module KPM ) is connected with residuals, and the output feature tensor of the regional perception module RPM is obtained by adding pixel by pixel. Re-enter the network backbone.

[0038]

[0039] Step 2: Combined with step 1, design a camouflaged target cascade perception network (CPNet).

[0040] CPNet sequentially cascades the feature extraction backbone Backbone, Feature Pyramid Networks (FPN), ternary cascade perception module and output detection head. The feature extraction backbone uses the Swin Transformer feature extraction module in the last four stages to downsample the input image to 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the original size, respectively. Then, the four sets of feature maps are input into the feature pyramid FPN for feature fusion. Let the high-level feature be F high , the low-level feature is F low , then after feature pyramid processing, the low-level features are processed as

[0041]

[0042] Among them, Conv(·, K = 1) represents a convolution layer with a convolution kernel size of 1, Conv(·, K = 3) represents a convolution layer with a convolution kernel size of 3, UP(·, 2X) represents an upsampling layer with a magnification factor of 2, and the processing method of the features of the highest layer of the backbone is After the feature pyramid is fused, the output is four three-dimensional tensors Based on this, the feature maps of four different scales are resized to F∈R through the convolution layer. 256×256×256 , and then superimpose them in the channel dimension to adjust the four three-dimensional tensors into a three-dimensional tensor F∈R 256×256×1024 , and then input into the ternary cascade perception module for important information perception, and finally output the target category, target box information and target confidence.

[0043] Step 3: Train model parameters and verify detection accuracy of the cascade perception network

[0044] A joint loss function is designed for the output prediction results. For the prediction of the target category, the focal loss (FocalLoss) function is used to obtain the target category loss value Loss_cls. For the prediction of the target box, the complete intersection over union (CIOU) function is used to obtain the target box loss value Loss_bbox. For the prediction of the target confidence, the cross entropy loss (Cross Entropy Loss) function is used to obtain the target confidence loss value Loss_conf. The total loss value Loss_all is:

[0045] Loss_all=Loss_cls+Loss_bbox+Loss_conf 05)

[0046] The target category loss value uses the focal loss function. Focal Loss can control the impact of positive and negative samples on the total loss value by adjusting the focal factor to obtain Loss_cls:

[0047] Loss_cls = -α(1-Y p ) γ log(Y p ) (16)

[0048] Among them, α is the focusing factor. Since the camouflaged target accounts for a small proportion in the image, α is selected as 0.25 to solve the problem of imbalance in the number of categories. γ represents the modulation factor and is selected as 2 to reduce the impact of a large number of high-confidence negative samples. Y l Represents the classification probability of different categories:

[0049]

[0050] The target box loss value uses the Complete Interp Over Union (CIOU) function. CIOU calculates the distance between the true target box and the predicted target box to get Loss_bbox:

[0051]

[0052] Where IOU represents the predicted target box bbox pred and the real target box bbox true The area intersection ratio between

[0053]

[0054] Represents the center point of the predicted target box The center point of the real target frame , β is the length of the farthest diagonal line between the predicted target box and the true target box, and δυ is used to measure the consistency of the aspect ratio between the predicted target box and the true target box:

[0055]

[0056]

[0057] Among them, w true 、h true Represents the width and height of the real target box, w pred 、h pred Represents the width and height of the prediction box. The target confidence loss value uses the cross entropy loss function to obtain Loss_conf:

[0058] Loss_conf=-log(y*conf+(1-y)*(1-conf)) (22)

[0059] Among them, y represents the true label, the background value is 0, the true target value is 1, and conf is the prediction confidence.

[0060] The evaluation indicators selected are the average precision (AP) and average recall (AR) with IOU between 0.5 and 0.95, as well as the average precision and average recall of small targets, medium targets and large targets distinguished by target size. Comprehensive evaluation indicators can better reflect the detection performance.

[0061] Compared with the prior art, the present invention has the following beneficial effects:

[0062] 1. In order to make up for the shortcomings of the existing camouflaged target detection technology, the present invention proposes a camouflaged target frame detection method based on ternary cascade perception, and designs a cascade perception network CPNet, which is composed of a feature extraction module, a ternary cascade perception module and a detection output head cascade. The detection output head outputs the target category, target frame and target confidence to achieve accurate frame detection of camouflaged targets.

[0063] 2. The present invention specifically designs a ternary cascade perception module TCPM, in which the feature perception module obtains the relative relationship information between features through a graph convolutional network, the key point perception module obtains the position information of the key points through average pooling of different dimensions, and the region perception module increases the receptive field by fusing different scales of dilated convolutions to obtain more region information; at the same time, the present invention designs a joint loss function for the output prediction results, and the total loss value is the sum of the target category loss value, the target box loss value and the target confidence loss value.

[0064] 3. The experimental results of the present invention can prove that the proposed cascade perception network CPNet has the best detection performance in terms of 10 comprehensive indicators due to its efficient use of feature relationship information, key point location information and regional information. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 It is a network flow chart of the cascade perception network CPNet proposed in the present invention.

[0066] Figure 2 It is a structural diagram of the feature extraction backbone Swin Transformer used in the present invention.

[0067] Figure 3 It is a structural diagram of the feature pyramid FPN proposed in the present invention.

[0068] Figure 4 It is a structural diagram of the ternary cascade perception module TCPM proposed in the present invention.

[0069] Figure 5 This is the visualization effect of the cascade perception network CPNet proposed in the present invention in identifying camouflaged targets.

[0070] in, Figure 1 In the ternary cascade perception module shown in the figure, the principle diagram is shown below the name of each submodule. Different shades of color represent different weight levels. The specific structure of each submodule is shown in Figure 4 . DETAILED DESCRIPTION

[0071] The present invention is further described in detail below in conjunction with the accompanying drawings. The embodiments described below are only used to illustrate the technical solution of the present invention, are part of the embodiments of the present invention, and are not intended to limit the scope of protection of the present invention.

[0072] The embodiment of the present invention discloses a visible light camouflage target frame detection method based on a cascade perception network CPNet, comprising the following steps:

[0073] Step 1: Ternary Cascade Perception Module TCPM is based on Figure 4 The structure diagram shown is constructed;

[0074] The modules are cascaded in sequence: feature perception module FPM, key point perception module KPM and region perception module RPM, and the output of the previous module is the input of the next module.

[0075] The feature perception module FPM obtains the relative relationship information between features through the graph convolutional network GCN. The attention operation of the feature perception module is ξ FPM (·), then the output of the feature perception module FPM is:

[0076]

[0077] The key point perception module KPM obtains the position information of the key points through the average pooling AP of different dimensions. The attention operation of the key point perception module is ξ KPM (·), then the output of the key point perception module KPM is:

[0078]

[0079] The regional perception module RPM increases the receptive field by fusing different scales of dilated convolution DConv to obtain more regional multi-scale information. The attention operation of the regional perception module is ξ RPM (·), then the output of the region perception module RPM is:

[0080]

[0081] Step 2: Cascade Perception Network CPNet is based on Figure 1 The network flow chart shown is constructed;

[0082] The cascade perception network CPNet is designed around the ternary cascade perception module TCPM, which cascades the feature extraction backbone Backbone, feature pyramid FPN, ternary cascade perception module and output detection head in sequence. The structure diagram of the feature extraction backbone SwinTransformer is shown in the figure. Figure 2 As shown, the structure of the feature pyramid FPN is as follows Figure 3 As shown in Figure 2, the feature pyramid FPN finally outputs four feature maps of different scales, and then the sizes are adjusted to F∈R through the convolution layer. 256×256×256 , superimpose them in the channel dimension and adjust the four three-dimensional tensors into a three-dimensional tensor F∈R 256×256×1024 , and then input into the ternary cascade perception module for important information perception, and finally the detection output head outputs the target category, target box information and target confidence.

[0083] Step 3: Construct a joint loss function to train the model parameters and verify the detection accuracy of the cascade perception network.

[0084] For the prediction of the target category, the focal loss function is used. Focal Loss can control the influence of positive and negative samples on the total loss value by adjusting the focal factor, and obtain the target category loss value Loss_cls. For the prediction of the target box, the complete intersection over union loss (CIOU) function is used. CIOU calculates the distance between the real target box and the predicted target box to obtain the target box loss value Loss_bbox. For the prediction of the target confidence, the cross entropy loss function is used to obtain the target confidence loss value Loss_conf. The total loss value Loss_all is the sum of the three loss values:

[0085] Loss_all=Loss_cls+Loss_bbox+Loss_conf (26)

[0086] The disguised target data set is input into the CPNet algorithm for parameter training and detection accuracy verification; the parameter training process is divided into three parts: data set processing, network hyperparameter selection and setting, and data set input into the CPNet algorithm for parameter update.

[0087] The dataset selects the COD10K general camouflaged target dataset released in 2020. The dataset contains 10,000 images, including 5,066 images with camouflaged targets, 3,000 background images, and 1,934 images without camouflaged targets. The dataset is divided into training set and validation set in a ratio of 6:4, which are used in the network model parameter training and algorithm verification stages respectively.

[0088] The hardware platform configuration used in the experimental training and testing phase is as follows: the graphics card is NVIDIA Quadro GV100, the video memory is 32G, and the Pytorch deep learning development framework is used. The CPU is Inter Xeon Silver 4210, and the operating system is Windows 10.

[0089] The evaluation indicators selected include the average precision (AP) and recall (AR) when IOU is between 0.5 and 0.95, as well as the accuracy and average recall of small targets, medium targets and large targets distinguished by target size. Comprehensive evaluation indicators can better reflect the detection performance.

[0090] The experimental results of the CPNet algorithm proposed in this invention and the excellent algorithms in recent years are compared as shown in Table 1. The data of other algorithms in Table 1 are all the results under the default parameters. The first line of each algorithm is the result of using a lightweight feature extraction backbone (such as R50, Tiny), and the second line is the result of using a deep feature extraction backbone (such as R101, Small).

[0091] The significance of the evaluation indicators in Table 1: AP 0.5-0.95 Refers to the average accuracy between IOU 0.5-0.95, AP 50 Refers to the accuracy when IOU is set to 0.5, AP 75 It refers to the accuracy when IOU is set to 0.75, AR1 refers to the recall calculated when only one prediction box with the maximum confidence is retained, AR5 refers to the recall calculated when five prediction boxes with the maximum confidence are retained, and AR 10 Refers to the recall rate calculated when retaining the prediction box of 1 degree, AP S and AR S Indicator representing small targets, AP M and AR M The indicator of the target in the representation, AP L and AR L Indicators that represent big goals.

[0092] Table 1 Comparison results of various algorithms

[0093]

[0094] It can be seen that the cascade perception network CPNet proposed in the present invention is better than other comparison algorithms in general. When using the lightweight feature extraction backbone, 8 indicators are better than other algorithms; when using the deep feature extraction backbone, 9 indicators are better than other algorithms, indicating that the cascade perception network CPNet has a strong pertinence for camouflaged target detection. Figure 5 Visualization of the cascaded perception network CPNet identifying some camouflaged targets.

Claims

1. A disguised target frame detection method based on ternary cascade perception, characterized by: The following steps are involved: Step 1: Construct a ternary cascade perception module TCPM (Ternary Cascade Perception Module) based on the characteristics of the camouflaged target; the ternary cascade perception module TCPM is constructed based on the characteristics of the camouflaged target, and specifically includes a feature perception module FPM (Featrue Perception Module), a key point perception module RPM (Key-point Perception Module, KPM) and a region perception module (Region Perception Module). The feature perception module obtains the relative relationship information between features through the graph convolution network GCN (Graph Convolution Network), and the key point perception module obtains the location information of the key points through average pooling AP (Average Pooling) of different dimensions. The region perception module increases the receptive field by fusing dilated convolution DConv (Dilated Convolution) of different scales to obtain more regional multi-scale information. The three perception modules are embedded in the backbone network in a cascade manner; Step 2: Combined with step 1, a cascade perception network CPNet (Cascade Perception Network) is designed. The cascade perception network CPNet is designed around the ternary cascade perception module TCPM, which cascades the feature extraction backbone Backbone, feature pyramid FPN (Feature Pyramid Networks,), ternary cascade perception module and output detection head in sequence. The feature extraction backbone Backbone uses the ST (Swin Transformer) feature extraction module in the last four stages to downsample the input image to 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the original size, respectively, and then inputs the four sets of feature maps into the feature perception module FPN for feature fusion. Let the high-level feature be The low-level features are After the feature pyramid FPN processing, the BR low-level feature processing is Among them, Conv(·, K = 1) represents a convolution layer with a convolution kernel size of 1, Conv(·, K = 3) represents a convolution layer with a convolution kernel size of 3, UP(·, 2X) represents an upsampling layer with a magnification factor of 2, and the processing method of the features of the highest layer of the backbone is After the feature pyramid is fused, the output is four three-dimensional tensors Based on this, the feature maps of four different scales are resized through the convolution layer. Then superimpose them in the channel dimension and adjust the four three-dimensional tensors into one three-dimensional tensor Then it is input into the ternary cascade perception module TCPM to perceive important information, and finally the detection output head outputs the target category, target box information and target confidence. Step 3: Construct a joint loss function to train the model parameters and verify the detection accuracy of the cascade perception network.

2. The method for detecting camouflaged target frames based on ternary cascade perception according to claim 1, characterized in that: The feature perception module FPM, for the input tensor First, the convolution layer combination CBR (Convolution + Batch Normalization + Activation Function Relu, abbreviated as CBR) is used for preprocessing to obtain the feature map Among them, Conv(BN(Relu(·))) represents the convolution layer combined CBR operation, and then the whole feature map is pooled in two dimensions through global average pooling (GAP) to generate a three-dimensional tensor pooling feature map. Among them, GAP(·, (W=1, H=1)) represents the operation of globally pooling the input feature map into a feature map with W=1 and H=1. In the convolutional neural network, the number of channels C represents the number of features extracted from the feature map, so the three-dimensional tensor It can be regarded as a set of 256 feature points. In order to further obtain the relative relationship information between features, the graph convolutional network GCN is used to learn the relationship between nodes and obtain a new three-dimensional tensor feature map with node relationships. Based on this, a Softmax normalization operation is performed on the channel dimension to normalize the relationship information to 0-1, and a three-dimensional tensor weight feature map is obtained. Among them, GCN(·) means using graph convolution network GCN to learn the relationship between nodes of three-dimensional tensor, Softmax(·, dim=C) means performing Softmax operation on channel dimension, and converting weight feature map With preprocessed feature map Multiply channel by channel to obtain the relation-weighted feature map Finally, a residual connection is performed with the original input, and pixel-by-pixel addition is performed to obtain the output feature tensor of the feature perception module FPM Re-enter the network backbone.

3. The method for detecting camouflaged target frames based on ternary cascade perception according to claim 1, characterized in that: The key point perception module KPM is the output tensor of the feature perception module FPM Use convolutional layers combined with CBR for preprocessing to obtain feature maps Unlike the feature perception module FPM, average pooling does not operate on the entire feature map, but divides it into two dimensions, W and H, to obtain the horizontal and vertical position information of the feature node in the feature map, and generate a pair of three-dimensional tensor pooling feature maps. and Among them, AP(·, (W=W, H=1)) represents the operation of pooling the input feature map into a feature map of W=W, H=1, and AP(·, (W=1, H=H)) represents the operation of pooling the input feature map into a feature map of W=1, H=H. The feature map containing the position information of the key points in different directions is multiplied to obtain the feature map with the global position information of the key points embedded. Then, the convolutional layer is combined with CBR to fuse the information between channels, and then the position information is normalized to 0-1 through the Softmax normalization operation of the channel dimension to obtain the weight feature map of the key point position information of the three-dimensional tensor. The weight feature map With preprocessed feature map Multiply channel by channel to obtain the relation-weighted feature map Finally, the output feature tensor of the input feature perception module FPM Perform residual connection and add pixel by pixel to get the output feature tensor of the key point perception module KPM Re-enter the network backbone.

4. The method for detecting camouflaged target frames based on ternary cascade perception according to claim 1, characterized in that: The region perception module RPM is the output tensor of the key point perception module Use convolutional layers combined with CBR for preprocessing to obtain feature maps Different from the key point perception module KPM, in order to obtain multi-scale information of key areas, it is necessary to perform feature extraction in a small range. Two dilated convolution DConv branches of different sizes are used to expand the receptive field and obtain multi-scale regional information. The dilated convolution kernel sizes are 3×3 and 5×5, respectively, and the dilation coefficients are both 2. Then the two sets of feature maps are superimposed to obtain a receptive field enhanced feature map with twice the number of channels. Among them, DConv(·, K=i, D=2) represents the dilated convolution with a dilation coefficient of 2 and a convolution kernel size of i. Add(·,·) represents the superposition operation of the two sets of feature maps in the channel dimension, and then the Softmax operation is performed on the channel dimension. Then, the convolution layer is combined with CBR to perform channel recovery and information fusion. Then, the position information is normalized to 0~1 through the Softmax normalization operation in the channel dimension, and the multi-scale information weight feature map of the three-dimensional tensor region is obtained. The weight feature map With preprocessed feature map Multiply channel by channel to obtain the relation-weighted feature map Finally, the output feature tensor of the key point perception module KPM input by the module Perform residual connection and add pixel by pixel to get the output feature tensor of the regional perception module RPM Re-enter the network backbone.

5. The method for detecting camouflaged target frames based on ternary cascade perception according to claim 1, characterized in that: The joint loss function in step 3 is a loss function of the training phase designed for the output prediction result. For the prediction of the target category, the focal loss (Focal Loss FL) function is used to obtain the target category loss value Loss_cls. For the prediction of the target box, the complete intersection over union (CIOU) loss function is used to obtain the target box loss value Loss_bbox. For the prediction of the target confidence, the cross entropy loss (Cross Entropy Loss CEL) function is used to obtain the target confidence loss value Loss_conf. The total loss value Loss_all is the sum of the three loss values. Loss_all=Loss_cls+Loss_bbox+Loss_conf (15) The target category loss value uses the focal loss function. Focal Loss can control the impact of positive and negative samples on the total loss value by adjusting the focal factor to obtain Loss_cls: Loss_cls=-α(1-Y p ) γ log(Y p ) (16) Among them, α is the focusing factor. Since the camouflaged target accounts for a small proportion in the image, α is selected as 0.25 to solve the problem of imbalance in the number of categories. γ represents the modulation factor and is selected as 2 to reduce the impact of a large number of high-confidence negative samples. Y P Represents the classification probability of different categories, y represents the true label, and the true target value is 1: The target box loss value uses the Complete Interp Over Union (CIOU) function. CIOU calculates the distance between the true target box and the predicted target box to get Loss_bbox: Where IOU represents the predicted target box bbox pred and the real target box bbox true The area intersection ratio between Represents the center point of the predicted target box The center point of the real target frame The Euclidean distance between them, β is the length of the farthest diagonal line between the predicted target frame and the true target frame, and δυ is used to measure the consistency of the aspect ratio between the predicted target frame and the true target frame. Among them, w true 、h true Represents the width and height of the real target box, w pred 、h pred Indicates the width and height of the prediction box; The target confidence loss value uses the cross entropy loss function, conf is the prediction confidence, and Loss_conf is obtained: Loss_conf=-log(y*conf+(1-y)*(1-conf)) (22).

6. The method for detecting camouflaged target frames based on ternary cascade perception according to claim 1, characterized in that: In the step 2, the detection accuracy is verified, and the evaluation indicators are the average accuracy (AP) and average recall (AR) of IOU between 0.5 and 0.95, as well as the average accuracy and average recall of small targets, medium targets and large targets distinguished by target size. Comprehensive evaluation indicators can better reflect the detection performance.

Images