A method and system for detecting small targets in infrared images

Abstract

The present invention relates to a method and system for detecting small targets in infrared images. The method for detecting small targets in infrared images based on a visual attention mechanism adopts a feature extraction network of the visual attention mechanism to extract features of infrared images, and a multi-scale feature fusion network splices and fuses deep and shallow feature maps to reduce information loss of small targets in infrared images, reduce false detection rate and missed detection rate of infrared small target detection, and improve network robustness.

Description

Technical Field

[0001] The present invention relates to the field of image target detection, and in particular to a method and system for detecting small infrared image targets. Background Art

[0002] Infrared small target detection technology is one of the important application technologies in the field of military target detection. The infrared radiation of the target in the infrared image has strong penetration, and the infrared radiation is not easily absorbed by the clouds, and is not easily affected by the weather and environment. The infrared detection system has the advantages of strong anti-interference ability, small size and easy installation, and can work all-weather in harsh environments. Therefore, infrared small target detection technology is widely used in the field of military target detection. For example, it is used in important tasks such as air defense early warning, missile interception, missile guidance, camouflage search, satellite remote sensing, etc.

[0003] At present, traditional infrared small target detection methods and deep learning methods have high accuracy and low missed detection rate for most infrared image scenes. However, infrared small targets lack characteristics such as color, contour, texture, etc. In the case of complex background and noise interference, the missed detection rate and false detection rate of existing infrared small target detection methods still need to be further reduced.

[0004] Therefore, how to design an infrared image weak target detection method and system that can address the high false detection rate and missed detection rate of infrared weak target detection in complex scenes such as complex backgrounds, light source interference, and noise interference, and reduce the false detection rate and missed detection rate of infrared weak target detection has become a technical problem that needs to be urgently solved in this field. Summary of the invention

[0005] The purpose of the present invention is to provide a method and system for detecting small infrared targets, by which the false detection rate and missed detection rate of small infrared target detection can be reduced.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A method for detecting small targets in infrared images, the method comprising the following steps:

[0008] Acquire sample infrared images;

[0009] Construct a visual attention mechanism feature extraction network;

[0010] Using the visual attention mechanism feature extraction network to extract features from the sample infrared image, and obtain feature maps of different scales;

[0011] Construct a multi-scale feature fusion network;

[0012] Using the multi-scale feature fusion network to concatenate and fuse the feature maps of different scales to obtain a feature map after multi-scale feature fusion;

[0013] The feature map after the multi-scale feature fusion is transmitted to the output end for screening, and the prediction result of the weak target in the infrared image is obtained;

[0014] Using a positioning loss function and a confidence loss function, the positioning loss and the confidence loss of the predicted result of the small target in the infrared image and the actual small target result in the manually annotated infrared image are calculated;

[0015] Based on the positioning loss and the confidence loss, the network parameters of the visual attention mechanism feature extraction network and the multi-scale feature fusion network are adjusted by backpropagation to obtain adjusted visual attention mechanism feature extraction network and multi-scale feature fusion network;

[0016] The positions of small targets in the target infrared image are predicted through the adjusted visual attention mechanism feature extraction network and multi-scale feature fusion network.

[0017] Optionally, the constructing of a visual attention mechanism feature extraction network specifically includes:

[0018] Combining the coordinate attention mechanism with the feature extraction network to obtain the visual attention mechanism feature extraction network;

[0019] The formula of the coordinate attention mechanism is shown in formula (1):

[0020] y c =x c ×σ(conv(f h ))×σ(conv(f w )) (1)

[0021] In formula (1), (f h ,f w )=Split(δ(F1((z h ,z w )))) represents the feature map of spatial information in the horizontal and vertical directions, where z h and z w They represent the feature maps after pooling in the horizontal and vertical coordinate directions, respectively. F1 represents the connection operation of the feature maps after pooling in the horizontal and vertical directions. δ represents the nonlinear activation function. Split represents the splitting operation along the horizontal and vertical dimensions. conv represents the convolution operation. σ represents the Sigmoid nonlinear activation function. c Represents the feature map output by the feature extraction network, y c Represents the feature map obtained after weighting by the coordinate attention mechanism.

[0022] Optionally, the using the visual attention mechanism feature extraction network to extract features from the infrared image to obtain feature maps of different scales specifically includes:

[0023] The feature map is extracted according to the following formula:

[0024]

[0025] In formula (2), X represents the input infrared image, conv i represents the i-th convolution operation for feature extraction of infrared images, x c represents the feature map extracted from the infrared image after multi-layer convolution operation; CA represents the weighted operation of the feature map by the coordinate attention mechanism, y c It represents the feature map obtained by weighting the coordinate attention mechanism, and n represents the number of layers of the feature extraction network.

[0026] Optionally, the multi-scale feature fusion network includes: an upsampling fusion layer and a downsampling fusion layer.

[0027] Optionally, the calculation formula of the upsampling fusion layer is as shown in formula (3):

[0028]

[0029] In formula (3), x i represents the i-th layer feature map of the visual attention mechanism feature extraction network, up represents the upsampling operation of the feature map, cat represents the concatenation of the feature map on the feature channel, conv i represents the i-th convolution operation after concatenating the feature maps for feature extraction, p i It represents the i-th layer feature map obtained by feature extraction of the upsampling fusion network, and n represents the number of layers of the upsampling fusion network.

[0030] Optionally, the calculation formula of the downsampling fusion layer is as shown in formula (4):

[0031]

[0032] In formula (4), p i represents the i-th layer feature map extracted by the upsampling fusion network, down represents the downsampling operation of the feature map, cat represents the splicing of the feature map on the feature channel, conv i represents the i-th convolution operation after concatenating the feature maps for feature extraction, f i It represents the i-th layer feature map obtained by feature extraction of the downsampling fusion network, and n represents the number of layers of the downsampling fusion network.

[0033] Optionally, the step of transmitting the feature map after the multi-scale feature fusion to an output end for screening to obtain a prediction result of a small target in the infrared image specifically includes:

[0034] Transmitting the feature map after the multi-scale feature fusion to an output end;

[0035] An anchor frame is used to predict the target position in the feature map after the multi-scale feature fusion and frame the target to obtain a target prediction frame;

[0036] The target prediction box is screened by using non-maximum suppression to obtain a screening result;

[0037] The prediction result of the small target in the infrared image is obtained according to the screening result.

[0038] Optionally, the formula of the positioning loss function is as shown in formula (5):

[0039]

[0040] In formula (5), L CIoU represents the positioning loss function, Measures the similarity of the aspect ratio of the real box and the predicted box, where w gt and h gt Respectively represent the width and height of the real box, w and h represent the width and height of the predicted box; To balance the loss caused by the aspect ratio and the loss caused by the IoU part, IoU represents the area of ​​the intersection of the real box and the predicted box, ρ 2 (b,b gt ) represents the square of the distance between the center point of the real box and the center point of the predicted box, c 2 Represents the square of the diagonal length of the rectangle enclosing the true box and the predicted box.

[0041] Optionally, the confidence loss function is expressed as follows:

[0042] FL=-(1-p t ) γ log(p t ) (6)

[0043] In formula (6), FL represents the confidence loss function, p t represents the probability of being predicted as the correct target, and γ represents the modulation coefficient.

[0044] The present invention also provides an infrared image small target detection system, which is characterized by comprising:

[0045] A sample infrared image acquisition module is used to acquire a sample infrared image;

[0046] The first network building module is used to build a visual attention mechanism feature extraction network;

[0047] A feature extraction module, used to extract features from the sample infrared image using the visual attention mechanism feature extraction network to obtain feature maps of different scales;

[0048] The second network construction module is used to construct a multi-scale feature fusion network;

[0049] A splicing and fusion module, used to splice and fuse the feature maps of different scales using the multi-scale feature fusion network to obtain a feature map after multi-scale feature fusion;

[0050] A screening module is used to transmit the feature map after the multi-scale feature fusion to the output end for screening, so as to obtain the prediction result of the weak target in the infrared image;

[0051] A calculation module, used to calculate the positioning loss and confidence loss of the predicted result of the small target in the infrared image and the actual small target result in the manually annotated infrared image by using the positioning loss function and the confidence loss function;

[0052] A network parameter adjustment module, used to adjust the network parameters of the visual attention mechanism feature extraction network and the multi-scale feature fusion network through reverse propagation based on the positioning loss and the confidence loss, so as to obtain adjusted visual attention mechanism feature extraction network and multi-scale feature fusion network;

[0053] The prediction module is used to predict the position of a small target in the target infrared image through the adjusted visual attention mechanism feature extraction network and the multi-scale feature fusion network.

[0054] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0055] The present invention provides a method and system for detecting small targets in infrared images. The method for detecting small targets in infrared images based on a visual attention mechanism adopts a feature extraction network of a visual attention mechanism to extract features of infrared images, and a multi-scale feature fusion network splices and fuses deep and shallow feature maps to reduce information loss of small targets in infrared images, reduce false detection rate and missed detection rate of infrared small target detection, and improve the robustness of the network. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0057] Figure 1 A flowchart of a method for detecting small targets in infrared images provided in Example 1 of the present invention;

[0058] Figure 2 It is a sample infrared image in the infrared weak target dataset;

[0059] Figure 3 This is the coordinate attention mechanism network structure diagram;

[0060] Figure 4 Extract network structure diagram for visual attention mechanism features;

[0061] Figure 5 This is a diagram of the multi-scale feature fusion network structure;

[0062] Figure 6 This is the prediction result diagram of infrared weak target;

[0063] Figure 7 This is a structural block diagram of an infrared image small target detection system provided in Example 2 of the present invention.

[0064] Explanation of symbols:

[0065] 1. Sample infrared image acquisition module; 2. First network construction module; 3. Feature extraction module; 4. Second network construction module; 5. Stitching and fusion module; 6. Screening module; 7. Calculation module; 8. Network parameter adjustment module; 9. Prediction module. DETAILED DESCRIPTION

[0066] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0067] The purpose of the present invention is to provide a method and system for detecting small infrared targets, by which the false detection rate and missed detection rate of small infrared target detection can be reduced.

[0068] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0069] Embodiment 1:

[0070] See also Figures 1 to 6 The present invention provides a method for detecting small targets in infrared images, the method comprising the following steps:

[0071] S1: Get the sample infrared image (such as Figure 2 shown);

[0072] S2: Construct a visual attention mechanism feature extraction network;

[0073] S3: extracting features from the sample infrared image using the visual attention mechanism feature extraction network to obtain feature maps of different scales;

[0074] S4: Construct a multi-scale feature fusion network;

[0075] S5: Using the multi-scale feature fusion network to concatenate and fuse the feature maps of different scales to obtain a feature map after multi-scale feature fusion;

[0076] S6: transmitting the feature map after the multi-scale feature fusion to the output end for screening, and obtaining the prediction result of the weak target in the infrared image;

[0077] S7: Calculate the positioning loss and confidence loss of the predicted result of the small target in the infrared image and the actual small target result in the manually annotated infrared image by using the positioning loss function and the confidence loss function;

[0078] S8: adjusting the network parameters of the visual attention mechanism feature extraction network and the multi-scale feature fusion network through reverse propagation based on the positioning loss and the confidence loss to obtain adjusted visual attention mechanism feature extraction network and multi-scale feature fusion network;

[0079] S9: Predict the position of small targets in the target infrared image through the adjusted visual attention mechanism feature extraction network and multi-scale feature fusion network.

[0080] In step S2, constructing a visual attention mechanism feature extraction network specifically includes:

[0081] The coordinate attention mechanism is combined with the feature extraction network to obtain the visual attention mechanism feature extraction network. Figure 3As shown in the figure, the position information is embedded into the channel attention, so that the feature map extracted by the feature network reduces useless background information while also having the coordinate information of the target.

[0082] The formula of the coordinate attention mechanism is shown in formula (1):

[0083] y c =x c ×σ(conv(f h ))×σ(conv(f w )) (1)

[0084] In formula (1), (f h ,f w )=Split(δ(F1((z h ,z w )))) represents the feature map of spatial information in the horizontal and vertical directions, where z h and z w They represent the feature maps after pooling in the horizontal and vertical coordinate directions, respectively. F1 represents the connection operation of the feature maps after pooling in the horizontal and vertical directions. δ represents the nonlinear activation function. Split represents the splitting operation along the horizontal and vertical dimensions. conv represents the convolution operation. σ represents the Sigmoid nonlinear activation function. c Represents the feature map output by the feature extraction network, y c Represents the feature map obtained after weighting by the coordinate attention mechanism.

[0085] In step S3, the feature extraction of the infrared image using the visual attention mechanism feature extraction network to obtain feature maps of different scales specifically includes:

[0086] The feature map is extracted according to the following formula:

[0087]

[0088] In formula (2), X represents the input infrared image, conv i represents the i-th convolution operation for feature extraction of infrared images, x c represents the feature map extracted from the infrared image after multi-layer convolution operation; CA represents the weighted operation of the feature map by the coordinate attention mechanism, y c represents the feature map obtained by weighting the coordinate attention mechanism, and n represents the number of layers of the feature extraction network. Among them, the network structure of the visual attention mechanism feature extraction network is as follows Figure 4 shown.

[0089] like Figure 5 As shown, the multi-scale feature fusion network includes: an upsampling fusion layer and a downsampling fusion layer.

[0090] The calculation formula of the upsampling fusion layer is shown in formula (3):

[0091]

[0092] In formula (3), x i represents the i-th layer feature map of the visual attention mechanism feature extraction network, up represents the upsampling operation of the feature map, cat represents the concatenation of the feature map on the feature channel, conv i represents the i-th convolution operation after concatenating the feature maps for feature extraction, p i It represents the i-th layer feature map obtained by feature extraction of the upsampling fusion network, and n represents the number of layers of the upsampling fusion network.

[0093] The calculation formula of the downsampling fusion layer is shown in formula (4):

[0094]

[0095] In formula (4), p i represents the i-th layer feature map extracted by the upsampling fusion network, down represents the downsampling operation of the feature map, cat represents the splicing of the feature map on the feature channel, conv i represents the i-th convolution operation after concatenating the feature maps for feature extraction, f i It represents the i-th layer feature map obtained by feature extraction of the downsampling fusion network, and n represents the number of layers of the downsampling fusion network.

[0096] In step S6, the feature map after the multi-scale feature fusion is transmitted to the output end for screening, and the prediction result of the small target in the infrared image is obtained, which specifically includes:

[0097] S61: Transmitting the feature map after the multi-scale feature fusion to an output end;

[0098] S62: using an anchor frame to predict the target position in the feature map after the multi-scale feature fusion and frame the target to obtain a target prediction frame;

[0099] S63: Using non-maximum suppression to filter the target prediction box to obtain a filtering result;

[0100] S64: Obtaining a prediction result of a small target in the infrared image according to the screening result.

[0101] In step S7, the formula of the positioning loss function is shown in formula (5):

[0102]

[0103] In formula (5), L CIoU represents the positioning loss function, Measures the similarity of the aspect ratio of the real box and the predicted box, where w gt and h gt Respectively represent the width and height of the real box, w and h represent the width and height of the predicted box; To balance the loss caused by the aspect ratio and the loss caused by the IoU part, IoU represents the area of ​​the intersection of the real box and the predicted box, ρ 2 (b,b gt ) represents the square of the distance between the center point of the real box and the center point of the predicted box, c 2 Represents the square of the diagonal length of the rectangle enclosing the true box and the predicted box.

[0104] The formula of the confidence loss function is shown in formula (6):

[0105] FL=-(1-p t ) γ log(p t ) (6)

[0106] In formula (6), FL represents the confidence loss function, p t represents the probability of being predicted as the correct target, and γ represents the modulation coefficient.

[0107] By training the visual attention mechanism feature extraction network and the multi-scale feature fusion network multiple times, the network after parameter training optimization can accurately predict the position of weak targets in the target infrared image (such as Figure 6 shown).

[0108] The invention discloses a method for detecting small targets in infrared images based on a visual attention mechanism. The method adopts a feature extraction network of the visual attention mechanism to extract features of infrared images. A multi-scale feature fusion network splices and fuses deep and shallow feature maps to reduce information loss of small targets in infrared images, reduce false detection rate and missed detection rate of infrared small targets, and improve the robustness of the network.

[0109] Embodiment 2:

[0110] See also Figure 7 The present invention provides an infrared image small target detection system, the system comprising:

[0111] The sample infrared image acquisition module 1 is used to acquire the sample infrared image;

[0112] The first network building module 2 is used to build a visual attention mechanism feature extraction network;

[0113] A feature extraction module 3 is used to extract features from the sample infrared image using the visual attention mechanism feature extraction network to obtain feature maps of different scales;

[0114] The second network construction module 4 is used to construct a multi-scale feature fusion network;

[0115] A splicing and fusion module 5 is used to splice and fuse the feature maps of different scales using the multi-scale feature fusion network to obtain a feature map after multi-scale feature fusion;

[0116] A screening module 6 is used to transmit the feature map after the multi-scale feature fusion to the output end for screening, and obtain the prediction result of the weak target in the infrared image;

[0117] A calculation module 7, used to calculate the positioning loss and confidence loss of the predicted result of the small target in the infrared image and the actual small target result in the manually annotated infrared image by using the positioning loss function and the confidence loss function;

[0118] A network parameter adjustment module 8, used to adjust the network parameters of the visual attention mechanism feature extraction network and the multi-scale feature fusion network through reverse propagation based on the positioning loss and the confidence loss, so as to obtain adjusted visual attention mechanism feature extraction network and multi-scale feature fusion network;

[0119] The prediction module 9 is used to predict the position of the weak target in the target infrared image through the adjusted visual attention mechanism feature extraction network and the multi-scale feature fusion network.

[0120] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0121] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A method for detecting small targets in infrared images, characterized in that: The following steps are involved: Acquire sample infrared images; Construct a visual attention mechanism feature extraction network; Using the visual attention mechanism feature extraction network to extract features from the sample infrared image, and obtain feature maps of different scales; Construct a multi-scale feature fusion network; Using the multi-scale feature fusion network to concatenate and fuse the feature maps of different scales to obtain a feature map after multi-scale feature fusion; The feature map after the multi-scale feature fusion is transmitted to the output end for screening, and the prediction result of the weak target in the infrared image is obtained; Using a positioning loss function and a confidence loss function, the positioning loss and the confidence loss of the predicted result of the small target in the infrared image and the actual small target result in the manually annotated infrared image are calculated; Based on the positioning loss and the confidence loss, the network parameters of the visual attention mechanism feature extraction network and the multi-scale feature fusion network are adjusted by backpropagation to obtain adjusted visual attention mechanism feature extraction network and multi-scale feature fusion network; Predicting the position of a small target in the target infrared image by using the adjusted visual attention mechanism feature extraction network and the multi-scale feature fusion network; The construction of the visual attention mechanism feature extraction network specifically includes: Combining the coordinate attention mechanism with the feature extraction network to obtain the visual attention mechanism feature extraction network; The formula of the coordinate attention mechanism is shown in formula (1): y c =x c ×σ(conv(f h ))×σ(conv(f w )) (1) In formula (1), (f h ,f w )=Split(δ(F1((z h ,z w )))) represents the feature map of spatial information in the horizontal and vertical directions, where z h and z w They represent the feature maps after pooling in the horizontal and vertical coordinate directions, respectively. F1 represents the connection operation of the feature maps after pooling in the horizontal and vertical directions. δ represents the nonlinear activation function. Split represents the splitting operation along the horizontal and vertical dimensions. conv represents the convolution operation. σ represents the Sigmoid nonlinear activation function. c Represents the feature map output by the feature extraction network, y c Represents the feature map obtained after weighting by the coordinate attention mechanism; The formula of the positioning loss function is shown in formula (5): In formula (5), L CIoU represents the positioning loss function, Measures the similarity of the aspect ratio of the real box and the predicted box, where w gt and h gt Respectively represent the width and height of the real box, w and h represent the width and height of the predicted box; To balance the loss caused by the aspect ratio and the loss caused by the IoU part, IoU represents the area of ​​the intersection of the real box and the predicted box, ρ 2 (b,b gt ) represents the square of the distance between the center point of the real box and the center point of the predicted box, c 2 Represents the square of the diagonal length of the rectangle that wraps the real box and the predicted box; The formula of the confidence loss function is shown in formula (6): FL=-(1-p t ) γ log(p t ) (6) In formula (6), FL represents the confidence loss function, p t represents the probability of being predicted as the correct target, and γ represents the modulation coefficient.

2. The infrared image small target detection method according to claim 1, characterized in that: The feature extraction network of the visual attention mechanism is used to extract features from the infrared image to obtain feature maps of different scales, specifically including: The feature map is extracted according to the following formula: In formula (2), X represents the input infrared image, conv i represents the i-th convolution operation for feature extraction of infrared images, x c represents the feature map extracted from the infrared image after multi-layer convolution operation; CA represents the weighted operation of the feature map by the coordinate attention mechanism, y c It represents the feature map obtained by weighting the coordinate attention mechanism, and n represents the number of layers of the feature extraction network.

3. The infrared image small target detection method according to claim 1, characterized in that: The multi-scale feature fusion network includes: an upsampling fusion layer and a downsampling fusion layer.

4. The infrared image small target detection method according to claim 3 is characterized in that: The calculation formula of the upsampling fusion layer is shown in formula (3): In formula (3), x i represents the i-th layer feature map of the visual attention mechanism feature extraction network, up represents the upsampling operation of the feature map, cat represents the concatenation of the feature map on the feature channel, conv i represents the i-th convolution operation after concatenating the feature maps for feature extraction, p i It represents the i-th layer feature map obtained by feature extraction of the upsampling fusion network, and n represents the number of layers of the upsampling fusion network.

5. The infrared image small target detection method according to claim 3, characterized in that: The calculation formula of the downsampling fusion layer is shown in formula (4): In formula (4), p i represents the i-th layer feature map extracted by the upsampling fusion network, down represents the downsampling operation of the feature map, cat represents the splicing of the feature map on the feature channel, conv i represents the i-th convolution operation after concatenating the feature maps for feature extraction, f i It represents the i-th layer feature map obtained by feature extraction of the downsampling fusion network, and n represents the number of layers of the downsampling fusion network.

6. The infrared image small target detection method according to claim 1, characterized in that: The step of transmitting the feature map after the fusion of the multi-scale features to the output end for screening to obtain the prediction result of the small target in the infrared image specifically includes: Transmitting the feature map after the multi-scale feature fusion to an output end; An anchor frame is used to predict the target position in the feature map after the multi-scale feature fusion and frame the target to obtain a target prediction frame; The target prediction box is screened by using non-maximum suppression to obtain a screening result; The prediction result of the small target in the infrared image is obtained according to the screening result.

7. An infrared image small target detection system based on the infrared image small target detection method according to claim 1, characterized in that: include: A sample infrared image acquisition module is used to acquire a sample infrared image; The first network building module is used to build a visual attention mechanism feature extraction network; A feature extraction module, used to extract features from the sample infrared image using the visual attention mechanism feature extraction network to obtain feature maps of different scales; The second network construction module is used to construct a multi-scale feature fusion network; A splicing and fusion module, used to splice and fuse the feature maps of different scales using the multi-scale feature fusion network to obtain a feature map after multi-scale feature fusion; A screening module is used to transmit the feature map after the multi-scale feature fusion to the output end for screening, so as to obtain the prediction result of the weak target in the infrared image; A calculation module, used to calculate the positioning loss and confidence loss of the predicted result of the small target in the infrared image and the actual small target result in the manually annotated infrared image by using the positioning loss function and the confidence loss function; A network parameter adjustment module, used to adjust the network parameters of the visual attention mechanism feature extraction network and the multi-scale feature fusion network through reverse propagation based on the positioning loss and the confidence loss, so as to obtain adjusted visual attention mechanism feature extraction network and multi-scale feature fusion network; The prediction module is used to predict the position of a small target in the target infrared image through the adjusted visual attention mechanism feature extraction network and the multi-scale feature fusion network.

Images