Infrared target detection method based on expanding receptive field

Abstract

The invention discloses an infrared target detection method based on expanding the receptive field, constructing a basic network structure for extracting features; inputting an image into the basic network structure, convolving through multiple convolutional layers layer by layer, and obtaining low-level feature maps and high-level feature maps of the image at different convolutional layers; expanding the receptive field by adding receptive field modules RFB‑d and RFB‑w designed and constructed after the low-level feature branch to obtain a low-level feature map with rich semantic information; respectively predicting the classification and bounding box to which each prediction box corresponding to each grid on the feature map belongs for the low-level feature map and the high-level feature map with rich semantic information; merging bounding boxes through a non-maximum suppression algorithm, and outputting the final prediction result of the entire network. The invention effectively improves the detection accuracy while improving the speed of the detection algorithm.

Description

Technical Field

[0001] The invention belongs to the field of infrared target detection, and in particular relates to an infrared target detection method based on expanding the receptive field. Background Art

[0002] As one of the basic problems of computer vision and digital image processing, target detection is the basis for the subsequent processing of many other computer vision tasks, such as instance segmentation, image understanding, target tracking, etc. The infrared system uses infrared detectors to sense the infrared radiation of objects to obtain information. It has the characteristics of all-weather operation, good concealment, easy penetration of smoke and dust, and strong anti-noise ability. At present, target detection in infrared scenes has important applications in autonomous driving, video surveillance, military and other fields.

[0003] Due to the lack of color and texture information in infrared images, low signal-to-noise ratio and contrast, severe background noise and low resolution, the traditional algorithm has a high false detection rate and poor robustness when using manually designed features to identify infrared targets. In recent years, the two-stage algorithm based on deep learning has gradually improved the detection accuracy, but the increase in computational complexity has also caused a decrease in detection speed to a certain extent. More and more practical application scenarios have higher and higher requirements for detection speed under the condition of ensuring a certain detection accuracy. The two-stage detection task can be decomposed into two single-stage detection tasks. The first stage task completes the candidate box generation, judges the background and target of the candidate box, and performs preliminary regression on the bounding box position. The second stage task realizes the final category prediction of the candidate box and the final position regression of the predicted box. The single-stage detection algorithm completes two tasks in a single stage, directly classifying and predicting the class to which the candidate box predicts the target, so it is of great significance to study the infrared target detection method based on the single stage. Summary of the invention

[0004] In view of this, the main purpose of the present invention is to provide an infrared target detection method based on expanding the receptive field.

[0005] To achieve the above object, the technical solution of the present invention is achieved as follows:

[0006] An embodiment of the present invention provides an infrared target detection method based on expanding the receptive field, the method comprising:

[0007] Construct a basic network structure for extracting features;

[0008] Input the image into the basic network structure, and convolve it layer by layer through multiple convolutional layers, and obtain low-level feature maps and high-level feature maps of the image at different convolutional layers;

[0009] The receptive field is expanded by adding the designed receptive field modules RFB-d and RFB-w after the low-level feature branches to obtain low-level feature maps with rich semantic information.

[0010] For the low-level feature map and the high-level feature map rich in semantic information, respectively predict the classification and bounding box to which each prediction box corresponding to each grid on the feature map belongs;

[0011] The bounding boxes are merged through the non-maximum suppression algorithm to output the final prediction result of the entire network.

[0012] In the above scheme, the basic network structure for extracting features is specifically: based on MobileNet, it is constructed by repeatedly stacking multiple different depth-wise separable convolution modules, the input image is downsampled through a 3×3 standard convolution layer, and then 13 depth-wise separable convolution modules DWS are stacked in the forward direction of the convolution. In the depth-wise separable convolution module, the depth convolution uses different convolution kernels for feature extraction for each channel of the input feature map, and point-by-point convolution is used to change the number of output channels to integrate information across channels.

[0013] In the above scheme, the image is input into the basic network structure, and convolved layer by layer through multiple convolutional layers, and the low-level feature map and high-level feature map of the image are obtained in different convolutional layers. Specifically, the output matrices of DWS11 and DWS13 obtained by processing the input image through the deep convolution module DWS are used as the low-level feature map of the image; and four additional convolutional layers C14, C15, C16, and C17 are directly added after the deep convolution layer DWS13, and the output of each of the four layers is used as the high-level feature map of the image.

[0014] In the above scheme, the receptive field is expanded by adding the designed and constructed receptive field modules RFB-d and RFB-w after the low-level feature branch to obtain a low-level feature map with rich semantic information. Specifically, the low-level feature map is processed by RFB-d to obtain the corresponding cascade of the output results of the four branches, and after splicing in the channel dimension, it is processed by the RFB-w module to obtain a low-level feature map with rich semantic information.

[0015] In the above scheme, the low-level feature map is processed by RFB-d to obtain the corresponding cascade of the output results of the four branches. Specifically, the RFB-d module has a total of four branch structures. For the input of the previous layer:

[0016] (1) Branch 1 generates output through a 1×1 convolutional layer;

[0017] (2) Branch 2 reduces the channel dimension through a 1×1 convolution layer, and adds a 3×3 dilated convolution with a dilation rate of 1 after the convolution layer;

[0018] (3) Branch 3 first passes through a 1×1 convolution layer to reduce the channel dimension, then decomposes the 3×3 convolution layer into cascaded 1×3 and 3×1 convolution operations, and adds a 3×3 dilated convolution with a dilation rate of 2;

[0019] (4) Branch 4 first passes through a 1×1 convolution layer to reduce the channel dimension, then decomposes the 5×5 convolution layer into two 3×3 cascaded convolution layers, and further decomposes them into 1×3 and 3×1 convolution layers, and adds a 3×3 dilated convolution with a dilation rate of 4;

[0020] The output results of the above four branches are cascaded accordingly, and finally spliced ​​in the channel dimension and input into the next RFB-w module.

[0021] In the above scheme, the RFB-w module is used to process and obtain a low-level feature map with rich semantic information. Specifically, the RFB-w module has a total of 4 branch structures. For the previous layer input:

[0022] (1) Branch 1 generates output through a 1×1 convolutional layer;

[0023] (2) Branch 2 passes through a 1×1 convolutional layer, and a 3×3 dilated convolution with a dilation rate of 1 is added after the convolutional layer;

[0024] (3) Branch 3 first passes through a 1×1 convolution layer to reduce the channel dimension, and then decomposes the 3×3 convolution layer into two parallel convolution branches of 1×3 and 3×1. A 3×3 dilated convolution with a dilation rate of 2 is added after each of the two convolution branches to generate two output results.

[0025] (4) Branch 4 first passes through a 1×1 convolution layer to reduce the channel dimension, then decomposes the 5×5 convolution layer into two 3×3 cascaded convolution layers, further decomposes one of them into two parallel convolution branches of 1×3 and 3×1, and adds a 3×3 dilated convolution with a dilation rate of 4 after each of the two convolution branches to expand the receptive field and generate two output results;

[0026] The six output results of the above four branches are cascaded accordingly and finally concatenated in the channel dimension as a low-level feature map with rich semantic information.

[0027] In the above scheme, the semantically rich low-level feature map and high-level feature map respectively predict the category and bounding box of each prediction box corresponding to each grid on the feature map, specifically: two low-level feature maps and four high-level feature maps processed by expanding the receptive field are respectively input into 6 detection branches, and each detection branch is divided into two 1×1 convolution branches, one branch outputs the category corresponding to each candidate box through the softmax layer, and the other branch outputs the corresponding bounding box through the linear regression layer. For each area, the classification and regression detection network will output the location and type of 6 possible targets. Among them, there are A target types, that is, the possibility of outputting A types of targets, which is related to the setting of the training set; the position parameter contains 4 data, including the center position coordinates, width, and height of the target bounding box.

[0028] In the above scheme, the method also includes training a basic network structure. When training the basic network structure, a designed and constructed scale-based intersection over union (SIOU) loss function is used, as follows:

[0029] SIOU=DIOU+βv

[0030] Among them, DIOU (Distance-IOU) is the distance loss function, β is the weight coefficient, and v is used to measure the similarity of the aspect ratio between the predicted box and the real box. The calculation formula is:

[0031]

[0032]

[0033] in, w and h are the width and height of the predicted box and the true box respectively.

[0034] The calculation of DIOU loss function is as follows

[0035]

[0036] First, calculate the intersection-and-union (IOU) of the predicted box A and the target real box B, then find the minimum bounding box C that contains both A and B, calculate the ratio of the area of ​​the remaining part of C that does not contain the product of A and B and the penalty term to the area of ​​C, and finally subtract it from the IOU.

[0037] Compared with the prior art, the present invention effectively improves the detection accuracy while improving the speed of the detection algorithm. The designed receptive field modules (RFB-w and RFB-d) expand the receptive field through dilated convolution, and extract and fuse features from multiple scales to obtain more sufficient semantic information, thereby improving the detection capability of small targets. The SIOU loss function calculates the similarity of the aspect ratio of the bounding box, adds the measure of the deformation of the bounding box, ensures the consistency of the scale of the target bounding box, effectively improves the problems of missed detection and bounding box deformation, and improves the detection capability of small targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The drawings described herein are used to disclose a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0039] Figure 1 An algorithm flow chart of an infrared target detection method based on expanding the receptive field is provided for an embodiment of the present invention;

[0040] Figure 2 A network structure diagram of an infrared target detection method based on expanding the receptive field is provided for an embodiment of the present invention;

[0041] Figure 3 A basic network structure diagram of an infrared target detection method based on expanding the receptive field is provided for an embodiment of the present invention;

[0042] Figure 4 (a) A structure diagram of an RFB-d receptive field module of an infrared target detection method based on expanding the receptive field is provided in an embodiment of the present invention;

[0043] Figure 4 (b) A structure diagram of an RFB-w receptive field module for an infrared target detection method based on expanding the receptive field according to an embodiment of the present invention;

[0044] Figure 5 A structural diagram of a multi-scale detection module of an infrared target detection method based on expanding the receptive field is provided for an embodiment of the present invention;

[0045] Figure 6 This is a partial detection result diagram of the algorithm provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0046] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0047] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, article or device. In the absence of more restrictions, an element defined by the sentence "includes a ..." does not exclude the existence of other identical elements in the process, article or device including the element.

[0048] The embodiment of the present invention provides an infrared target detection method based on expanding the receptive field, such as Figure 1 As shown, the method is:

[0049] Step 1: Build a basic network structure for extracting features;

[0050] Specifically, Figure 2 , 3 As shown in the figure, based on MobileNet, it is constructed by repeatedly stacking multiple different depth-wise separable convolution modules. The input image is downsampled through a 3×3 standard convolution layer, and then 13 depth-wise separable convolution modules DWS are stacked in the forward direction of the convolution. In the depth-wise separable convolution module, the depth-wise convolution uses different convolution kernels for feature extraction for each channel of the input feature map, and point-by-point convolution is used to change the number of output channels to integrate information across channels.

[0051] Step 2: Input the image into the basic network structure, and convolve it layer by layer through multiple convolutional layers to obtain low-level feature maps and high-level feature maps of the image at different convolutional layers;

[0052] Specifically, the output matrices of DWS11 and DWS13 obtained by processing the input image through the deep convolution module DWS are used as the low-level feature map of the image; four additional convolution layers C14, C15, C16, and C17 are directly added after the deep convolution layer DWS13, and the output of each of the four layers is used as the high-level feature map of the image.

[0053] Step 3: By adding the designed receptive field modules RFB-d and RFB-w after the low-level feature branches, the receptive field is expanded to obtain low-level feature maps with rich semantic information;

[0054] Specifically, Figure 4 As shown in the figure, the low-level feature map is processed by RFB-d to obtain the corresponding cascade of the output results of the four branches, and then spliced ​​in the channel dimension and processed by the RFB-w module to obtain a low-level feature map with rich semantic information.

[0055] The low-level feature map is processed by RFB-d to obtain the corresponding cascade of the output results of the four branches. Specifically, the RFB-d module has a total of four branch structures. For the previous layer input:

[0056] (1) Branch 1 generates output through a 1×1 convolutional layer;

[0057] (2) Branch 2 reduces the channel dimension through a 1×1 convolution layer, and adds a 3×3 dilated convolution with a dilation rate of 1 after the convolution layer;

[0058] (3) Branch 3 first passes through a 1×1 convolution layer to reduce the channel dimension, then decomposes the 3×3 convolution layer into cascaded 1×3 and 3×1 convolution operations, and adds a 3×3 dilated convolution with a dilation rate of 2;

[0059] (4) Branch 4 first passes through a 1×1 convolution layer to reduce the channel dimension, then decomposes the 5×5 convolution layer into two 3×3 cascaded convolution layers, and further decomposes them into 1×3 and 3×1 convolution layers, and adds a 3×3 dilated convolution with a dilation rate of 4;

[0060] The output results of the above four branches are cascaded accordingly, and finally spliced ​​in the channel dimension and input into the next RFB-w module.

[0061] The RFB-w module will be used to process the low-level feature maps with rich semantic information. Specifically, the RFB-w module has a total of 4 branch structures. For the previous layer input:

[0062] (1) Branch 1 generates output through a 1×1 convolutional layer;

[0063] (2) Branch 2 passes through a 1×1 convolutional layer, and a 3×3 dilated convolution with a dilation rate of 1 is added after the convolutional layer;

[0064] (3) Branch 3 first passes through a 1×1 convolution layer to reduce the channel dimension, and then decomposes the 3×3 convolution layer into two parallel convolution branches of 1×3 and 3×1. A 3×3 dilated convolution with a dilation rate of 2 is added after each of the two convolution branches to generate two output results.

[0065] (4) Branch 4 first passes through a 1×1 convolution layer to reduce the channel dimension, then decomposes the 5×5 convolution layer into two 3×3 cascaded convolution layers, further decomposes one of them into two parallel convolution branches of 1×3 and 3×1, and adds a 3×3 dilated convolution with a dilation rate of 4 after each of the two convolution branches to expand the receptive field and generate two output results;

[0066] The six output results of the above four branches are cascaded accordingly and finally concatenated in the channel dimension as a low-level feature map with richer semantic information.

[0067] Step 4: respectively predicting the classification and bounding box to which each prediction box corresponding to each grid on the feature map belongs for the low-level feature map and the high-level feature map rich in semantic information;

[0068] Specifically, two low-level feature maps and four high-level feature maps with expanded receptive fields are input into six detection branches respectively, and each detection branch is divided into two 1×1 convolution branches. One branch outputs the category corresponding to each candidate box through the softmax layer, and the other branch outputs the corresponding bounding box through the linear regression layer. For each area, the classification and regression detection network will output the location and type of six possible targets. Among them, there are A target types, that is, the possibility of outputting A types of targets, which is related to the setting of the training set; the position parameter contains 4 data, including the center position coordinates, width, and height of the target bounding box.

[0069] like Figure 5 As shown in the figure, multi-scale detection is performed, and the feature maps are input into 6 detection branches respectively. Each detection branch is divided into two 1×1 convolution branches. One branch outputs the category corresponding to each candidate box through the softmax layer, and the other branch outputs the corresponding bounding box through the linear regression layer.

[0070] For the a priori boxes generated by feature maps of different layers, assuming that m layers of feature maps are used for prediction, the size of the a priori boxes in each layer of feature maps is

[0071]

[0072] Among them, S k is the ratio of the prior box in the feature map of the kth layer, S max and S min are the ratios of the top and bottom prior frames in the feature map of layer m, respectively. For the different aspect ratios of the prior frames of the feature map of the same layer, the width and height of each prior frame need to be calculated. If a is set r = 5 different aspect ratios such as {1, 2, 3, 1 / 2, 1 / 3}, then the width of each prior box is

[0073]

[0074] Similarly, the height of each prior box is

[0075]

[0076] The center coordinates of the prior box are as follows

[0077]

[0078] Among them, |f k| is the value of the edge length of the kth feature map, i,j∈[0,|f k |), 0.5 is the offset. Generally, when the aspect ratio is 1, an additional size is added. The prior box.

[0079] Each center point of the feature map will generate 6 prior boxes of different sizes. The prior boxes of different sizes and aspect ratios in the feature map cover almost all the positions to be detected and can complete the detection of all target areas of the input. By calculating the score of each prior box for each category and background, the category to which the prior box belongs is determined.

[0080] Since reducing the size of the convolution kernel has almost no effect on the detection accuracy but can reduce the amount of calculation, a 1×1 convolution kernel is used in the detection module, and the number of prior frames is set to 3, 6, 6, 6, 6, 6. Assuming that the input image resolution is 300×300, the target detection method with expanded receptive field outputs feature maps of 6 scales, 19×19, 10×10, 5×5, 3×3, 2×2 and 1×1. The first layer of the network predicts 3 bounding boxes per grid, and the other 5 layers of the network predict 6 bounding boxes per grid. Each bounding box has 4 parameters (the horizontal and vertical coordinates of the center point of the prior frame, the length and width of the prior frame), and a total of 5 categories are predicted. Finally, the number of generated prior frames is 19×19×3+(10×10+5×5+3×3+2×2+1×1)×6=1917.

[0081] Step 5: Merge the bounding boxes through the non-maximum suppression algorithm and output the final prediction result of the entire network.

[0082] like Figure 6 As shown in the figure, it can be seen that the algorithm can accurately detect multi-scale targets including small targets without missing any targets. Figure 6 (a) Pedestrians at medium and long distances Figure 6 (b) Cars at different distances, and Figure 6 (b) When the distance between the bicycle and the smaller bicycle is long. Figure 6 (d) shows a very small puppy. In addition, the bounding box of the object is located very accurately and does not deform.

[0083] Basic network structure and parameter settings:

[0084] 1. When training the network, use the scale-based intersection-over-union (SIOU) loss function, expressed as

[0085] SIOU=DIOU+βv

[0086] Among them, DIOU is the distance loss function, β is the weight coefficient, and v is used to measure the similarity of the aspect ratio between the predicted box and the real box. The calculation formula is:

[0087]

[0088]

[0089] in, w and h are the width and height of the predicted box and the true box respectively.

[0090] The calculation of DIOU loss function is as follows

[0091]

[0092] First, the intersection-and-union (IOU) of the predicted box A and the target real box B is calculated, then the minimum bounding box C that contains both A and B is found, the ratio of the area of ​​the remaining part of C that does not contain the product of A and B and the penalty term to the area of ​​C is calculated, and finally subtracted from the IOU. Finally, according to the loss function calculation result, the weights are updated using the stochastic gradient descent method with momentum.

[0093] 2. The convolution kernel sizes of Conv14 to Conv17 are all 1×1. The size, step size, and whether the output feature map is input into the detection branch are shown in Table 1.1.

[0094] Table 1.1 Some convolutional layer parameter settings

[0095]

[0096] 3. The infrared target detection method based on expanding the receptive field extracts features from 6 effective feature maps of different scales for detection. These 6 detection modules include two convolution branches of size 1×1, one of which performs classification prediction on the target and the other performs bounding box regression on the target. The input size, number of bounding boxes, and number of classification branch and regression branch channels of each detection module are shown in Table 1.2.

[0097] Table 1.2 Parameter settings for each detection module

[0098]

[0099] 4. Some hyperparameters in the experiment are set as follows: scaling the original image to 512×512, the number of batches is 64, the momentum is 0.9, the decay is 0.0005, the initial learning rate is 0.001, the maximum number of iterations is 50,000, the learning strategy adopts multistep, and the learning rate is decayed tenfold at iterations 40,000 and 45,000.

[0100] The above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention.

Claims

1. An infrared target detection method based on expanding the receptive field, characterized in that: The method is: Construct a basic network structure for extracting features; Input the image into the basic network structure, and convolve it layer by layer through multiple convolutional layers, and obtain low-level feature maps and high-level feature maps of the image at different convolutional layers; The receptive field is expanded by adding the designed receptive field modules RFB-d and RFB-w after the low-level feature branches to obtain low-level feature maps with rich semantic information. For the low-level feature map and the high-level feature map rich in semantic information, respectively predict the classification and bounding box to which each prediction box corresponding to each grid on the feature map belongs; The bounding boxes are merged through the non-maximum suppression algorithm to output the final prediction result of the entire network; The basic network structure for extracting features is constructed as follows: based on MobileNet, it is constructed by repeatedly stacking multiple different depth-wise separable convolution modules, the input image is downsampled through a 3×3 standard convolution layer, and then 13 depth-wise separable convolution modules DWS are stacked in the forward direction of the convolution. In the depth-wise separable convolution module, the depth convolution uses different convolution kernels for each channel of the input feature map to extract features, and point-by-point convolution is used to change the number of output channels to integrate information across channels; The image is input into the basic network structure, and convolved layer by layer through multiple convolution layers, and the low-level feature map and high-level feature map of the image are obtained at different convolution layers. Specifically, the output matrices of DWS11 and DWS13 obtained by processing the input image through the deep convolution module DWS are used as the low-level feature map of the image; and four additional convolution layers C14, C15, C16, and C17 are directly added after the deep convolution layer DWS13, wherein the output of each of the four layers is used as the high-level feature map of the image; The receptive field is expanded by adding the designed and constructed receptive field modules RFB-d and RFB-w after the low-level feature branch to obtain a low-level feature map with rich semantic information. Specifically, the low-level feature map is processed by RFB-d to obtain the corresponding cascade of the output results of the four branches, and then spliced ​​in the channel dimension and processed by the RFB-w module to obtain a low-level feature map with rich semantic information. The low-level feature map is processed by RFB-d to obtain the corresponding cascade of the output results of the four branches. Specifically, the RFB-d module has a total of four branch structures. For the previous layer input: (1) Branch 1 generates output through a 1×1 convolutional layer; (2) Branch 2 reduces the channel dimension through a 1×1 convolution layer, and adds a 3×3 dilated convolution with a dilation rate of 1 after the convolution layer; (3) Branch 3 first passes through a 1×1 convolution layer to reduce the channel dimension, then decomposes the 3×3 convolution layer into cascaded 1×3 and 3×1 convolution operations, and adds a 3×3 dilated convolution with a dilation rate of 2; (4) Branch 4 first passes through a 1×1 convolution layer to reduce the channel dimension, then decomposes the 5×5 convolution layer into two 3×3 cascaded convolution layers, and further decomposes them into 1×3 and 3×1 convolution layers, and adds a 3×3 dilated convolution with a dilation rate of 4; The output results of the above four branches are cascaded accordingly, and finally spliced ​​in the channel dimension and input into the next RFB-w module.

2. The infrared target detection method based on expanding the receptive field according to claim 1 is characterized in that: The RFB-w module will be used to process the low-level feature maps with rich semantic information. Specifically, the RFB-w module has a total of 4 branch structures. For the previous layer input: (1) Branch 1 generates output through a 1×1 convolutional layer; (2) Branch 2 passes through a 1×1 convolutional layer, and a 3×3 dilated convolution with a dilation rate of 1 is added after the convolutional layer; (3) Branch 3 first passes through a 1×1 convolution layer to reduce the channel dimension, and then decomposes the 3×3 convolution layer into two parallel convolution branches of 1×3 and 3×1. A 3×3 dilated convolution with a dilation rate of 2 is added after each of the two convolution branches to generate two output results. (4) Branch 4 first passes through a 1×1 convolution layer to reduce the channel dimension, then decomposes the 5×5 convolution layer into two 3×3 cascaded convolution layers, further decomposes one of them into two parallel convolution branches of 1×3 and 3×1, and adds a 3×3 dilated convolution with a dilation rate of 4 after each of the two convolution branches to expand the receptive field and generate two output results; The six output results of the above four branches are cascaded accordingly and finally concatenated in the channel dimension as a low-level feature map with rich semantic information.

3. The infrared target detection method based on expanding the receptive field according to claim 2 is characterized in that: The low-level feature map and high-level feature map with rich semantic information respectively predict the category and bounding box of each prediction box corresponding to each grid on the feature map, specifically: two low-level feature maps processed with expanded receptive fields and four high-level feature maps are respectively input into 6 detection branches, each detection branch is divided into two 1×1 convolution branches, one branch outputs the category corresponding to each candidate box through the softmax layer, and the other branch outputs the corresponding bounding box through the linear regression layer. For each area, the classification and regression detection network will output the positions and types of 6 possible targets; among them, there are A target types in total, that is, the possibility of outputting A types of targets is related to the setting of the training set; the position parameter contains 4 data, including the center position coordinates, width, and height of the target bounding box.

4. The infrared target detection method based on expanding the receptive field according to claim 3 is characterized in that: The method also includes training a basic network structure. When training the basic network structure, a designed and constructed scale-based intersection over union (SIOU) loss function is used, as follows: SIOU=DIOU+βv Among them, DIOU (Distance-IOU) is the distance loss function, β is the weight coefficient, and v is used to measure the similarity of the aspect ratio between the predicted box and the real box. The calculation formula is: in, w and h are the width and height of the predicted box and the real box respectively; The calculation of DIOU loss function is as follows First, calculate the intersection-and-union (IOU) of the predicted box A and the target real box B, then find the minimum bounding box C that contains both A and B, calculate the ratio of the area of ​​the remaining part of C that does not contain the product of A and B and the penalty term to the area of ​​C, and finally subtract it from the IOU.

Images