Weak and small heat source target detection method under space background

Abstract

The invention provides a weak and small heat source target detection method under a space background, and the method comprises the steps: obtaining a plurality of space background infrared images with real heat source labels, inputting the space background infrared images into a constructed weak and small heat source target detection network model, and obtaining a heat source detection result in the space background infrared images, the weak and small heat source target detection network model comprises an image enhancement unit, an encoder, a feature fusion unit and a decoder; establishing a total loss function based on the heat source detection result and the real heat source label, optimizing the constructed weak and small heat source target detection network model, and obtaining a trained weak and small heat source target detection network model when the total loss function value reaches the minimum value; and inputting a space background infrared image to be detected into the trained weak and small heat source target detection network model to obtain a heat source detection result in the space background infrared image to be detected. The method can improve the detection precision of the weak and small heat source target under the space background, and avoids wrong detection and missing detection.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and more specifically, to a method for detecting weak heat source targets in a space background. Background Art

[0002] In recent years, with the increase in the number of my country's in-orbit spacecraft and the continuous strengthening of its space power, my country's in-orbit spacecraft are also facing some potential threats, such as being illegally detected by other countries' spacecraft or being hit by space debris. Therefore, it is particularly important to ensure that in-orbit spacecraft can detect and avoid these abnormal targets in a timely manner.

[0003] In space scenes, small and weak targets to be detected usually have the following characteristics. First, the pixels occupied by these targets on the imaging system are very small, with almost no obvious shape, texture and color features; second, the targets are imaged at extremely long distances, and are affected by cosmic rays and particle disturbances during the imaging process, resulting in smaller imaging pixel values ​​of the targets relative to other noise (such as stars), and lower signal-to-noise ratio (SNR); at the same time, the targets are similar to other noise (such as stars, background radiation) in visual effects, making them difficult to distinguish, and the targets may be located at the far end of the orbit, with small radiation themselves, and appear dim visually.

[0004] Based on the above characteristics of infrared weak small targets under the space background, the method is required to accurately extract the features of weak small targets and retain small targets in the deep features, otherwise the small targets may be lost. In addition, it is necessary to be able to distinguish the target from the rest of the visually similar noise, otherwise misdetection may occur. The prior art discloses an infrared small target detection method and system to solve the technical problems of poor effect and large computational complexity of the existing small target detection algorithm, but the method focuses on global and local feature extraction, and does not set jump connections between different deep features. Multiple downsampling easily leads to the loss of small targets in deep features. In addition, the method does not perform image enhancement on the image to be detected under the space background. Since the infrared target under the space background is darker and difficult to distinguish from the surrounding environment than ordinary infrared small targets, if the original image is not enhanced, the network may not be able to distinguish between the target, background and noise, resulting in poor detection effect. Summary of the invention

[0005] In order to overcome the defects that the heat source target in the space background is small in size and has weak characteristics, and is difficult to distinguish from the surrounding environment, resulting in false detection and missed detection, the present invention provides a method for detecting weak heat source targets in the space background.

[0006] In order to solve the above technical problems, the technical solution of the present invention is as follows:

[0007] The present invention proposes a method for detecting weak heat source targets in a space background, comprising:

[0008] Obtain several space background infrared images with real heat source annotations;

[0009] Inputting the space background infrared image with real heat source annotations into the constructed weak heat source target detection network model to obtain the heat source detection result in the space background infrared image;

[0010] A total loss function is established based on the heat source detection results and the real heat source annotations, and the constructed weak heat source target detection network model is optimized. When the total loss function value reaches the minimum, a trained weak heat source target detection network model is obtained;

[0011] Obtain the space background infrared image to be detected, input it into the trained weak heat source target detection network model, and obtain the heat source detection result in the space background infrared image to be detected.

[0012] Preferably, the weak heat source target detection network model includes an image enhancement unit, a first encoder, a second encoder, a third encoder, a fourth encoder, a fifth encoder, a sixth encoder, a first feature fusion unit, a second feature fusion unit, a third feature fusion unit, a feature embedding unit, a first decoder, a second decoder, a detection unit, a first feature splicing layer, a second feature splicing layer, a third feature splicing layer, a fourth feature splicing layer and a fifth feature splicing layer;

[0013] The input end of the image enhancement unit is connected to the input end of the first encoder, the output end of the image enhancement unit is connected to the input end of the second encoder, the output end of the second encoder is respectively connected to the second input end of the first feature fusion unit and the input end of the second feature splicing layer, the output end of the first encoder is respectively linked to the first input end of the first feature fusion unit and the input end of the first feature splicing layer, the first and second output ends of the first feature fusion unit are respectively connected to the input end of the first feature splicing layer and the input end of the second feature splicing layer, the output end of the first feature splicing layer is connected to the input end of the third encoder, and the output end of the second feature splicing layer is connected to the input end of the fourth encoder;

[0014] The output end of the third encoder is connected to the first input end of the second feature fusion unit and the input end of the third feature splicing layer respectively, the output end of the fourth encoder is connected to the second input end of the second feature fusion unit and the input end of the fourth feature splicing layer respectively, the first and second output ends of the second feature fusion unit are connected to the input end of the third feature splicing layer and the input end of the fourth feature splicing layer respectively, the output end of the third feature splicing layer is connected to the input end of the fifth encoder, and the output end of the fourth feature splicing layer is connected to the input end of the sixth encoder;

[0015] The output end of the fifth encoder is connected to the first input end of the third feature fusion unit, the output end of the sixth encoder is connected to the second input end of the second feature fusion unit, the first and second output ends of the third feature fusion unit are respectively connected to the input end of the fifth feature splicing layer, the output end of the fifth feature splicing layer is connected to the input end of the feature embedding unit, and the feature embedding unit, the first decoder, the second decoder and the detection unit are connected in sequence.

[0016] Preferably, the image enhancement unit includes a first Mamba module, a second Mamba module, and a first MconvBlock module, a second MconvBlock module, a third MconvBlock module, a first Concat layer, a first DWconv layer, a first RConvBlock module, a second RConvBlock module, a third RConvBlock module and a first FConvBlock module connected in sequence;

[0017] The input end of the first MconvBlock module is used as the input end of the image enhancement unit, the input end of the first Mamba module is connected to the input end of the first MconvBlock module, the output end of the first Mamba module is connected to the input end of the second Mamba module, and the output end of the second Mamba module is connected to the input end of the first Concat layer.

[0018] The output ends of the first RConvBlock module and the second RConvBlock module are also connected to the input end of the first FConvBlock module; the output end of the first FConvBlock module serves as the output end of the image enhancement unit.

[0019] Preferably, the first MconvBlock module, the second MconvBlock module, and the third MconvBlock module have the same structure, including a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a fifth convolutional layer, a sixth convolutional layer, a seventh convolutional layer, a first average pooling layer, a sixth feature concatenation layer, a first batch of normalization layers, and a first ReLU activation function;

[0020] The input end of the first convolutional layer is used as the input end of the MconvBlock module, the output end of the first convolutional layer is connected to the input ends of the second convolutional layer, the third convolutional layer, the fourth convolutional layer and the first average pooling layer, the output end of the third convolutional layer is connected to the input end of the fifth convolutional layer, the output end of the fourth convolutional layer is connected to the input end of the sixth convolutional layer, the input end of the first average pooling layer is connected to the input end of the seventh convolutional layer, and the output ends of the second convolutional layer, the fifth convolutional layer, the sixth convolutional layer and the seventh convolutional layer are all connected to the input end of the sixth feature splicing layer;

[0021] The sixth feature concatenation layer, the first batch of normalization layers and the first ReLU activation function are connected in sequence; the output end of the first ReLU activation function serves as the output end of the MconvBlock module.

[0022] Preferably, the FConvBlock module comprises a second Concat layer, a second DwConv layer, a second ReLU activation function, a second batch normalization layer, a third DwConv layer, a third ReLU activation function, a third batch normalization layer and a third Concat layer connected in sequence;

[0023] The output of the second DwConv layer and the output of the second batch normalization layer are also connected to the input of the third Concat layer;

[0024] The input end of the second Concat layer is used as the input end of the FConvBlock module, and the output end of the third Concat layer is used as the output end of the FConvBlock module.

[0025] Preferably, the encoder includes a fourth RconvBlock module, a fifth RconvBlock module, a sixth RconvBlock module, a seventh RconvBlock module, a first AConvBlock module and a second AConvBlock module;

[0026] The fourth RconvBlock module, the fifth RconvBlock module and the first AConvBlock module are connected in sequence, and the output end of the fourth RconvBlock module is also connected to the input end of the first AConvBlock module and the input end of the sixth RconvBlock module;

[0027] The output end of the sixth RconvBlock module is connected to the input end of the seventh RconvBlock module and the input end of the second AConvBlock module respectively, the output end of the seventh RconvBlock module is connected to the input end of the second AConvBlock module, and the output end of the second AConvBlock module is connected to the input end of the first AConvBlock module;

[0028] The input end of the fourth RconvBlock module is used as the input end of the encoder, and the output end of the first AConvBlock module is used as the output end of the encoder;

[0029] The AConvBlock module includes a fourth Concat layer, an eighth convolutional layer, a channel space attention submodule, a fourth DwConv layer, a fourth ReLU activation function and a fourth batch normalization layer connected in sequence;

[0030] The input end of the fourth Concat layer is used as the input end of the AConvBlock module, and the output end of the fourth batch normalization layer is used as the output end of the AConvBlock module.

[0031] Preferably, the decoder comprises a ninth convolutional layer, a first linear projection submodule, a first normalization layer, a first normalization layer, a third Mamba module, a first feature addition layer, a second normalization layer, a first multilayer perceptron and a second feature addition layer connected in sequence;

[0032] The output end of the third Mamba module is connected to the first input end of the first feature addition layer, and the input end of the first multilayer perceptron is connected to the first input end of the second feature addition layer;

[0033] The output end of the first normalization layer is connected to the second input end of the first feature addition layer, and the output end of the first feature addition layer is also connected to the second input end of the second feature addition layer;

[0034] The input end of the ninth convolutional layer is used as the input end of the decoder, and the output end of the second feature addition layer is used as the output end of the decoder.

[0035] Preferably, the feature fusion unit includes an eighth RconvBlock module, a ninth RconvBlock module, a tenth RconvBlock module, an eleventh RconvBlock module, a twelfth RconvBlock module, a thirteenth RconvBlock module, a fourteenth RconvBlock module, a fifteenth RconvBlock module, a sixteenth RconvBlock module, a seventeenth RconvBlock module, a third feature addition layer, a fourth feature addition layer, a fifth feature addition layer, a sixth feature addition layer, a seventh feature addition layer, an eighth feature addition layer, a first feature multiplication layer, a second feature multiplication layer, a feature subtraction layer, a global average pooling layer, a fifth Concat layer, a sixth Concat layer, a tenth convolution layer and an eleventh convolution layer;

[0036] The input end of the eighth RconvBlock module is used as the first input end of the feature fusion unit, and the input end of the eighth RconvBlock module is connected to the first input end of the first feature multiplication layer and the input end of the thirteenth RconvBlock module respectively;

[0037] The input end of the fifteenth RconvBlock module serves as the second input end of the feature fusion unit, and the input end of the fifteenth RconvBlock module is respectively connected to the first input end of the second feature multiplication layer and the input end of the tenth RconvBlock module;

[0038] The eighth RconvBlock module, the ninth RconvBlock module and the third feature addition layer are connected in sequence, the output end of the ninth RconvBlock module is connected to the first input end of the third feature addition layer, the output end of the eighth RconvBlock module is also connected to the second input end of the third feature addition layer, and the output end of the third feature addition layer is connected to the first input end of the feature subtraction layer;

[0039] The tenth RconvBlock module, the eleventh RconvBlock module and the fourth feature addition layer are connected in sequence, the output end of the eleventh RconvBlock module is connected to the first input end of the fourth feature addition layer, the output end of the tenth RconvBlock module is also connected to the second input end of the fourth feature addition layer and the second input end of the fifth feature addition layer, and the output end of the fourth feature addition layer is connected to the second input end of the feature subtraction layer;

[0040] The thirteenth RconvBlock module, the fourteenth RconvBlock module and the fifth feature addition layer are connected in sequence, the output end of the fourteenth RconvBlock module is connected to the first input end of the fifth feature addition layer, the output end of the thirteenth RconvBlock module is also connected to the second input end of the sixth feature addition layer, and the output end of the fifth feature addition layer is connected to the first input end of the seventh feature addition layer;

[0041] The fifteenth RconvBlock module, the sixteenth RconvBlock module and the sixth feature addition layer are connected in sequence, the output end of the sixteenth RconvBlock module is connected to the first input end of the sixth feature addition layer, and the output end of the sixth feature addition layer is connected to the second input end of the seventh feature addition layer;

[0042] The output end of the feature subtraction layer is connected to the input end of the twelfth RconvBlock module, and the output end of the twelfth RconvBlock module is connected to the first input end of the eighth feature addition layer;

[0043] The output end of the seventh feature addition layer is connected to the input end of the seventeenth RconvBlock module, and the output end of the seventeenth RconvBlock module is connected to the second input end of the eighth feature addition layer;

[0044] The output end of the eighth feature addition layer is connected to the input end of the fifth Concat layer, the input end of the sixth Concat layer and the input end of the global average pooling layer respectively, and the output end of the global average pooling layer is connected to the second input end of the first feature multiplication layer and the second input end of the fourth two-feature multiplication layer respectively;

[0045] The first feature multiplication layer, the fifth Concat layer and the tenth convolution layer are connected in sequence; the second feature multiplication layer, the sixth Concat layer and the eleventh convolution layer are connected in sequence; the output end of the tenth convolution layer and the output end of the eleventh convolution layer serve as the first and second output ends of the feature fusion unit respectively.

[0046] Preferably, the first RConvBlock module, the second RConvBlock module, the third RConvBlock module, the fourth RConvBlock module, the fifth RConvBlock module, the sixth RConvBlock module, the seventh RConvBlock module, the eighth RConvBlock module, the ninth RConvBlock module, the tenth RConvBlock module, the eleventh RConvBlock module, the twelfth RConvBlock module, the thirteenth RConvBlock module, the fourteenth RConvBlock module, the fifteenth RConvBlock module, the sixteenth RConvBlock module and the seventeenth RConvBlock module have the same structure, including a fifth DwConv layer, a fifth ReLU activation function, a fifth batch normalization layer, a sixth DwConv layer, a sixth ReLU activation function, a sixth batch normalization layer and a seventh Concat layer connected in sequence;

[0047] The output of the fifth DwConv layer and the output of the fifth batch normalization layer are also connected to the input of the seventh Concat layer;

[0048] The input end of the fifth DwConv serves as the input end of each of the RConvBlock modules, and the output end of the seventh Concat layer serves as the output end of each of the RConvBlock modules.

[0049] Preferably, the first Mamba module, the second Mamba module and the third Mamba module have the same structure, including a second linear projection submodule, a third linear projection submodule, a fourth linear projection submodule, a twelfth convolutional layer, a first SiLU activation function, a second SiLU activation function, a state space submodule and a second normalization layer;

[0050] The second linear projection submodule, the twelfth convolutional layer, the first SiLU activation function, the state space submodule, the second normalization layer and the fourth linear projection submodule are connected in sequence;

[0051] The input end of the second linear projection submodule is used as the input end of each of the Mamba modules, the input end of the third linear projection submodule is connected to the input end of the first linear projection submodule, the output end of the third linear projection submodule is connected to the input end of the second SiLU activation function, and the output end of the second SiLU activation function is connected to the input end of the second normalization layer;

[0052] The output end of the fourth linear projection submodule serves as the output end of each of the Mamba modules.

[0053] The present invention also proposes a weak heat source target detection system under a space background, which is used to implement the above-mentioned weak heat source target detection method under a space background, comprising:

[0054] A data acquisition module is used to acquire several infrared images of space background with real heat source annotations;

[0055] The first detection module is used to input the space background infrared image with real heat source annotations into the constructed weak heat source target detection network model to obtain the heat source detection result in the space background infrared image;

[0056] A model training module is used to establish a total loss function based on the heat source detection results and the real heat source annotations, optimize the constructed weak heat source target detection network model, and obtain a trained weak heat source target detection network model when the total loss function value reaches the minimum;

[0057] The second detection module is used to obtain the space background infrared image to be detected, input it into the trained weak heat source target detection network model, and obtain the heat source detection result in the space background infrared image to be detected.

[0058] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0059] The present invention proposes a method for detecting weak and small heat source targets under a space background, obtaining several space background infrared images with real heat source annotations; inputting the space background infrared images with real heat source annotations into a constructed weak and small heat source target detection network model to obtain heat source detection results in the space background infrared images, wherein the weak and small heat source target detection network model includes an image enhancement unit, an encoder, a feature fusion unit, and a decoder; establishing a total loss function based on the heat source detection results and the real heat source annotations, optimizing the constructed weak and small heat source target detection network model, and obtaining a trained weak and small heat source target detection network model when the total loss function value reaches the minimum; obtaining a space background infrared image to be detected, inputting it into the trained weak and small heat source target detection network model, and obtaining heat source detection results in the space background infrared image to be detected. The present invention can improve the detection accuracy of weak and small heat source targets under a space background and avoid false detection and missed detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 This is a schematic diagram of the structure of the weak heat source target detection network model described in Example 2;

[0061] Figure 2 This is a schematic diagram of the structure of the image enhancement unit described in Example 2;

[0062] Figure 3 This is a schematic diagram of the structure of the FConvBlock module described in Example 2;

[0063] Figure 4 This is a schematic diagram of the structure of the MconvBlock module described in Example 2;

[0064] Figure 5 This is a schematic diagram of the structure of the encoder described in Example 2;

[0065] Figure 6 This is a schematic diagram of the structure of the channel space attention submodule described in Example 2;

[0066] Figure 7 This is a schematic diagram of the structure of the feature fusion unit described in Example 2;

[0067] Figure 8 This is a schematic diagram of the structure of the RConvBlock module described in Example 2;

[0068] Fig. 9 This is a schematic diagram of the structure of the decoder described in Example 2;

[0069] Fig.10 This is a schematic diagram of the structure of the Mamba module described in Example 2;

[0070] Fig.11This is a schematic diagram of the structure of the weak heat source target detection system under the space background described in Example 3. DETAILED DESCRIPTION

[0071] The drawings are for illustrative purposes only and should not be construed as limiting the present patent;

[0072] In order to better illustrate the present embodiment, some parts in the drawings may be omitted, enlarged or reduced, and do not represent the size of the actual product;

[0073] It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0074] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0075] Example 1

[0076] This embodiment provides a method for detecting a weak heat source target in a space background, including:

[0077] Obtain several space background infrared images with real heat source annotations;

[0078] Inputting the space background infrared image with real heat source annotations into the constructed weak heat source target detection network model to obtain the heat source detection result in the space background infrared image;

[0079] A total loss function is established based on the heat source detection results and the real heat source annotations, and the constructed weak heat source target detection network model is optimized. When the total loss function value reaches the minimum, a trained weak heat source target detection network model is obtained;

[0080] Obtain the space background infrared image to be detected, input it into the trained weak heat source target detection network model, and obtain the heat source detection result in the space background infrared image to be detected.

[0081] In the specific implementation process, the space background infrared image with real heat source annotations is first input into the weak heat source target detection network model for training, and then the space background infrared image to be detected is obtained and input into the trained weak heat source target detection network model to obtain the heat source detection result in the space background infrared image to be detected.

[0082] Example 2

[0083] This embodiment provides a method for detecting a weak heat source target in a space background, including:

[0084] Obtain several space background infrared images with real heat source annotations;

[0085] Inputting the space background infrared image with real heat source annotations into the constructed weak heat source target detection network model to obtain the heat source detection result in the space background infrared image;

[0086] A total loss function is established based on the heat source detection results and the real heat source annotations, and the constructed weak heat source target detection network model is optimized. When the total loss function value reaches the minimum, a trained weak heat source target detection network model is obtained;

[0087] Obtain the space background infrared image to be detected, input it into the trained weak heat source target detection network model, and obtain the heat source detection result in the space background infrared image to be detected.

[0088] like Figure 1 As shown, it is a structural schematic diagram of a weak heat source target detection network model, and the weak heat source target detection network model includes an image enhancement unit, a first encoder, a second encoder, a third encoder, a fourth encoder, a fifth encoder, a sixth encoder, a first feature fusion unit, a second feature fusion unit, a third feature fusion unit, a feature embedding unit, a first decoder, a second decoder, a detection unit, a first feature splicing layer, a second feature splicing layer, a third feature splicing layer, a fourth feature splicing layer and a fifth feature splicing layer;

[0089] The input end of the image enhancement unit is connected to the input end of the first encoder, the output end of the image enhancement unit is connected to the input end of the second encoder, the output end of the second encoder is respectively connected to the second input end of the first feature fusion unit and the input end of the second feature splicing layer, the output end of the first encoder is respectively linked to the first input end of the first feature fusion unit and the input end of the first feature splicing layer, the first and second output ends of the first feature fusion unit are respectively connected to the input end of the first feature splicing layer and the input end of the second feature splicing layer, the output end of the first feature splicing layer is connected to the input end of the third encoder, and the output end of the second feature splicing layer is connected to the input end of the fourth encoder;

[0090] The output end of the third encoder is connected to the first input end of the second feature fusion unit and the input end of the third feature splicing layer respectively, the output end of the fourth encoder is connected to the second input end of the second feature fusion unit and the input end of the fourth feature splicing layer respectively, the first and second output ends of the second feature fusion unit are connected to the input end of the third feature splicing layer and the input end of the fourth feature splicing layer respectively, the output end of the third feature splicing layer is connected to the input end of the fifth encoder, and the output end of the fourth feature splicing layer is connected to the input end of the sixth encoder;

[0091] The output end of the fifth encoder is connected to the first input end of the third feature fusion unit, the output end of the sixth encoder is connected to the second input end of the second feature fusion unit, the first and second output ends of the third feature fusion unit are respectively connected to the input end of the fifth feature splicing layer, the output end of the fifth feature splicing layer is connected to the input end of the feature embedding unit, and the feature embedding unit, the first decoder, the second decoder and the detection unit are connected in sequence.

[0092] (1) Image enhancement unit

[0093] Because in the space environment, weak infrared targets are usually dim in the space background and difficult to distinguish from the background. The objects in the target and scene are of different sizes and shapes, and the positions are mostly non-fixed. Single-scale feature extraction may lose some important information and affect the detection effect. Through image enhancement, the contrast between the target and the background and other noise can be significantly improved, making the target more obvious and easier to be recognized by the detection model. Therefore, in order to solve this problem, this embodiment designs an image enhancement unit. The image enhancement unit draws on the idea of ​​the Inception structure and includes a multi-scale feature unit MconvBlock, that is, the MconvBlock module, which uses convolution kernels of different sizes to obtain receptive fields of different ranges, thereby more comprehensively capturing local target and scene feature information and enhancing the network's adaptability to multiple scales.

[0094] like Figure 2 , which is a schematic diagram of the structure of an image enhancement unit, wherein the image enhancement unit includes a first Mamba module, a second Mamba module, and a first MconvBlock module, a second MconvBlock module, a third MconvBlock module, a first Concat layer, a first DWconv layer, a first RConvBlock module, a second RConvBlock module, a third RConvBlock module, and a first FConvBlock module connected in sequence;

[0095] The input end of the first MconvBlock module is used as the input end of the image enhancement unit, the input end of the first Mamba module is connected to the input end of the first MconvBlock module, the output end of the first Mamba module is connected to the input end of the second Mamba module, and the output end of the second Mamba module is connected to the input end of the first Concat layer.

[0096] The output ends of the first RConvBlock module and the second RConvBlock module are also connected to the input end of the first FConvBlock module; the output end of the first FConvBlock module serves as the output end of the image enhancement unit.

[0097] like Figure 3 , which is a schematic diagram of the structure of the FConvBlock module, wherein the FConvBlock module includes a second Concat layer, a second DwConv layer, a second ReLU activation function, a second batch normalization layer, a third DwConv layer, a third ReLU activation function, a third batch normalization layer and a third Concat layer connected in sequence;

[0098] The output end of the second DwConv layer and the output end of the second batch of normalization layers are also connected to the input end of the third Concat layer; the input end of the second Concat layer serves as the input end of the FConvBlock module, and the output end of the third Concat layer serves as the output end of the FConvBlock module.

[0099] like Figure 4 , which is a schematic diagram of the structure of the MconvBlock module, wherein the first MconvBlock module, the second MconvBlock module, and the third MconvBlock module have the same structure, including a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a fifth convolutional layer, a sixth convolutional layer, a seventh convolutional layer, a first average pooling layer, a sixth feature concatenation layer, a first batch of normalization layers, and a first ReLU activation function;

[0100] The input end of the first convolutional layer is used as the input end of the MconvBlock module, the output end of the first convolutional layer is connected to the input ends of the second convolutional layer, the third convolutional layer, the fourth convolutional layer and the first average pooling layer, the output end of the third convolutional layer is connected to the input end of the fifth convolutional layer, the output end of the fourth convolutional layer is connected to the input end of the sixth convolutional layer, the input end of the first average pooling layer is connected to the input end of the seventh convolutional layer, and the output ends of the second convolutional layer, the fifth convolutional layer, the sixth convolutional layer and the seventh convolutional layer are all connected to the input end of the sixth feature splicing layer;

[0101] The sixth feature concatenation layer, the first batch of normalization layers and the first ReLU activation function are connected in sequence; the output end of the first ReLU activation function serves as the output end of the MconvBlock module.

[0102] In order to reduce computational consumption and improve network performance, the MconvBlock module mainly uses small convolution kernels of 1×1×1 and 3×3×3. The 1×1×1 convolution is used to achieve dimensionality reduction of the number of channels and reduce the amount of calculation. At the same time, it combines the high correlation features between different channels at the same spatial position to achieve information fusion between channels. Convolution kernels of other sizes, such as 3×3×3 and 5×5×5 convolutions, ensure a larger receptive field and feature diversity. In addition, the global information is also supplemented by the Mamba module. By combining local and global information, the module can obtain richer image details. Among them, the state space submodule of the Mamba module (also called the state space model, SSM) can be expressed by a linear ordinary differential equation with evolution parameters:

[0103] h′(t)=Ah(t)+Bx(t)

[0104] y(t)=Ch(t)+Dx(t)

[0105] Among them, x(t) represents the input sequence, h(t) represents the latent state, h′(t) represents the update of the latent state, and y(t) represents the predicted output sequence. The state space submodule maps x(t) to the response y(t) through the latent space h(t). Then the two multi-scale feature extraction units in the lower branch and the two Mamba modules in the upper branch form a dual-branch multi-scale feature extraction path to extract local and global features.

[0106] The MconvBlock module consists of four convolutional layers of different sizes, namely 1×1×1, 3×3×3, 5×5×5 convolutional layers and an average pooling layer. The first convolutional layer is 3×3×3×128, that is, 3×3×3 convolution kernels, 128 channels, the second convolutional layer is 1×1×1×64, the third convolutional layer is 1×1×1×32, the fourth convolutional layer is 1×1×1×32, the fifth convolutional layer is 3×3×3×32, the sixth convolutional layer is 5×5×5×32, and the seventh convolutional layer is 3×3×3×32. The first branch uses 1×1×1 convolution to focus on small-sized objects, the second branch uses 3×3×3 convolution to focus on medium-sized objects, and the third branch uses 5×5×5 convolution to focus on large-sized objects. The average pooling layer of the fourth branch is used to change the feature arrangement and reduce the thickness of the feature map. After executing the above steps, the feature information of each scale is fused through the sixth feature concatenation layer and then normalized and processed by the first ReL u Activation function processing to form a new feature map.

[0107] In the image enhancement unit, the image is first input into three parallel multi-scale feature extraction units and two Mamba modules. After the multi-scale feature extraction is completed, the feature map is input into the image restoration unit. This module consists of 3 RconvBlocks and multiple jump connections. Each RconvBlock contains a combination of a normal convolution layer, a grouped convolution layer, a ReLU activation function, and a batch normalization layer. Grouped convolution is used here to reduce the amount of calculation and promote network lightweight. In addition, each RconvBlock contains two groups of jump connections, which splice the input features with each group of output features to retain more original image details and semantic information. The input of each RconvBlock is spliced ​​by shallow jump connections. The input of the nth RconvBlock is defined as:

[0108]

[0109] Among them, Concat[·] represents the concatenation operation, and F(·) represents the 3×3 convolution operation. Through the direct transmission of features at different stages, the jump connection helps to fuse feature information at different levels, enhance the network's ability to capture local details, and retain more original image details and semantic information. After three RconvBlocks and jump connections, the image enhancement unit outputs an enhanced infrared image.

[0110] To optimize the image enhancement effect, this section introduces the following loss function:

[0111]

[0112] Among them, y i represents the enhanced image output by the image enhancement module, y′ i Represents the ideal enhanced image (i.e., groundtruth). By constraining the output of the enhanced image through this loss function, the model can gradually approach the ideal image enhancement effect during the training process.

[0113] (2) Encoder

[0114] It is difficult to fully capture faint and inconspicuous targets by simply relying on the original image for feature extraction. Although the previously obtained enhanced image improves the contrast and brightness of the target area, it may also introduce noise or lose the original details. This makes it easy for the single-path feature extraction to miss key details and affect the detection accuracy. Therefore, after being processed by the image enhancement module, the original image and the enhanced image are simultaneously sent to the dual-branch parallel encoder for parallel feature extraction.

[0115] like Figure 5, which is a schematic diagram of the structure of an encoder, wherein the encoder includes a fourth RconvBlock module, a fifth RconvBlock module, a sixth RconvBlock module, a seventh RconvBlock module, a first AConvBlock module and a second AConvBlock module;

[0116] The fourth RconvBlock module, the fifth RconvBlock module and the first AConvBlock module are connected in sequence, and the output end of the fourth RconvBlock module is also connected to the input end of the first AConvBlock module and the input end of the sixth RconvBlock module;

[0117] The output end of the sixth RconvBlock module is connected to the input end of the seventh RconvBlock module and the input end of the second AConvBlock module respectively, the output end of the seventh RconvBlock module is connected to the input end of the second AConvBlock module, and the output end of the second AConvBlock module is connected to the input end of the first AConvBlock module;

[0118] The input end of the fourth RconvBlock module is used as the input end of the encoder, and the output end of the first AConvBlock module is used as the output end of the encoder; the AConvBlock module includes a fourth Concat layer, an eighth convolutional layer, a channel space attention submodule, a fourth DwConv layer, a fourth ReLU activation function and a fourth batch normalization layer connected in sequence; the input end of the fourth Concat layer is used as the input end of the AConvBlock module, and the output end of the fourth batch normalization layer is used as the output end of the AConvBlock module.

[0119] The encoder uses a U-shaped network structure as the basic network structure, and is composed of three layers of networks. Among them, the first layer consists of two RconvBlock modules and an AConvBlock module, namely the fourth RconvBlock module, the fifth RconvBlock module and the first AConvBlock module, the second layer consists of an RconvBlock and an AConvBlock, namely the sixth RconvBlock module and the second AConvBlock module, and the last layer consists of only one RconvBlock, namely the seventh RconvBlock module. The RconvBlock module of each layer is the same as the RconvBlock module in the image enhancement unit, and will not be repeated here. The last Block of the first and second layers is an AConvBlock module, which receives the jump connection from the same layer and the lower layer as input. After receiving the feature map transmitted by the jump connection, they will be spliced ​​in the channel dimension, and then the channel space attention submodule is used to perform feature fusion of different dimensions to reduce the semantic gap, and finally the output is obtained through the convolution layer + Relu activation function + batch normalization layer. Here, grouped convolution is also used to reduce the amount of calculation and promote the lightweight of the network.

[0120] like Figure 6 , which is a schematic diagram of the structure of the channel space attention submodule, the channel space attention submodule includes a first maximum pooling layer, a second maximum pooling layer, a second average pooling layer, a third average pooling layer, a second multi-layer perceptron, a first fully connected layer, a second fully connected layer, a ninth feature addition layer, a tenth feature addition layer, a first Sigmoid activation function, a second Sigmoid activation function, a third feature multiplication layer, a fourth feature multiplication layer, and a thirteenth convolutional layer;

[0121] The first input end of the third feature multiplication layer is the input end of the channel space attention submodule, the input end of the first maximum pooling layer and the input end of the second average pooling layer are connected to the first input end of the first feature multiplication layer, the output end of the first maximum pooling layer and the output end of the second average pooling layer are both connected to the input end of the second multilayer perceptron, the output end of the second multilayer perceptron is respectively connected to the input ends of the first fully connected layer and the second fully connected layer, the output ends of the first fully connected layer and the second fully connected layer are respectively connected to the first and second input ends of the ninth feature addition layer, the output end of the ninth feature addition layer is connected to the input end of the first Sigmoid activation function, and the input end of the first Sigmoid activation function is connected to the second input end of the third feature multiplication layer;

[0122] The output end of the third feature multiplication layer is connected to the second input end of the fourth feature multiplication layer, the input end of the second maximum pooling layer, and the input end of the third average pooling layer respectively, the output end of the second maximum pooling layer and the output end of the third average pooling layer are connected to the first and second input ends of the tenth feature addition layer respectively, the output end of the tenth feature addition layer is connected to the input end of the thirteenth convolutional layer, the output end of the third convolutional layer is connected to the input end of the second Sigmoid activation function, and the output end of the second Sigmoid activation function is connected to the first input end of the second feature multiplication layer;

[0123] The output of the fourth feature multiplication layer serves as the output of the channel space attention submodule.

[0124] The feature map first enters the channel space attention submodule, and the feature map first enters the channel attention submodule, and the input feature map is compressed through the average pooling layer and the maximum pooling layer to generate the average pooling feature and the maximum pooling feature X max , and then X arg and X max Input to a shared weight layer consisting of two fully connected layers. The features output by the shared weight layer are fused and then activated by the Sigmoid function to generate the final channel attention M C ∈R nchannels×1×1×1 , that is, the first Sigmoid activation function outputs the channel attention M C In addition, in order to reduce computational consumption, the activation size of the hidden layer is set to nchannels×1×1×1. In summary, the overall formula for channel attention is:

[0125] M C =σ(FC(MaxPool(X))+FC(AvgPool(X)))

[0126] Among them, X represents the input feature, MaxPool represents the maximum pooling layer, AvgPool represents the average pooling layer, FC represents the fully connected layer, and σ represents the Sigmoid activation function. Spatial attention can be regarded as a supplement to channel attention. After obtaining the channel attention, the channel attention and input features are input into the third feature multiplication layer for fusion, and the output feature X is obtained. c , as the input feature of the spatial attention submodule. Spatial attention submodule X C The input can be described by the following formula:

[0127]

[0128] in, Represents the element-wise multiplication operation of the feature multiplication layer. In order to make full use of channel information, the three-dimensional features are generated through average pooling and maximum pooling operations. Figure X avg and X max Then they are fused and activated through a standard convolutional layer and Sigmoid function to generate a three-dimensional space attention, called M S The spatial attention calculation formula is:

[0129] M S =σ(F 3×3×3 ([MaxPo0l(X C ); AvgPool(X C )]))

[0130] Among them, F 3×3×3 is a three-dimensional convolution with a convolution kernel size of 3×3×3. C and M S Fusion, to obtain the output of the entire channel space attention submodule, can be expressed by the following formula:

[0131]

[0132] After the above process, the model extracts multi-scale feature maps for subsequent detection processing.

[0133] (3) Feature fusion unit

[0134] Since the enhanced image often deviates from the original image in details, the features extracted by the dual-branch parallel encoder often have misalignment problems, which directly affects the effect of target detection. To solve this problem, a feature fusion module is designed to align the feature maps on the two branches and combine their respective advantages to improve the comprehensiveness and expression ability of the overall features. The original image retains richer details, while the enhanced image has higher contrast and brightness. Fusion of the features of these two types of images can better utilize their respective advantages and improve the overall performance of the model.

[0135] like Figure 7As shown, it is a structural schematic diagram of a feature fusion unit, which includes an eighth RconvBlock module, a ninth RconvBlock module, a tenth RconvBlock module, an eleventh RconvBlock module, a twelfth RconvBlock module, a thirteenth RconvBlock module, a fourteenth RconvBlock module, a fifteenth RconvBlock module, a sixteenth RconvBlock module, a seventeenth RconvBlock module, a third feature addition layer, a fourth feature addition layer, a fifth feature addition layer, a sixth feature addition layer, a seventh feature addition layer, an eighth feature addition layer, a first feature multiplication layer, a second feature multiplication layer, a feature subtraction layer, a global average pooling layer, a fifth Concat layer, a sixth Concat layer, a tenth convolution layer and an eleventh convolution layer;

[0136] The input end of the eighth RconvBlock module is used as the first input end of the feature fusion unit, and the input end of the eighth RconvBlock module is connected to the first input end of the first feature multiplication layer and the input end of the thirteenth RconvBlock module respectively;

[0137] The input end of the fifteenth RconvBlock module serves as the second input end of the feature fusion unit, and the input end of the fifteenth RconvBlock module is respectively connected to the first input end of the second feature multiplication layer and the input end of the tenth RconvBlock module;

[0138] The eighth RconvBlock module, the ninth RconvBlock module and the third feature addition layer are connected in sequence, the output end of the ninth RconvBlock module is connected to the first input end of the third feature addition layer, the output end of the eighth RconvBlock module is also connected to the second input end of the third feature addition layer, and the output end of the third feature addition layer is connected to the first input end of the feature subtraction layer;

[0139] The tenth RconvBlock module, the eleventh RconvBlock module and the fourth feature addition layer are connected in sequence, the output end of the eleventh RconvBlock module is connected to the first input end of the fourth feature addition layer, the output end of the tenth RconvBlock module is respectively connected to the second input end of the fourth feature addition layer and the second input end of the fifth feature addition layer, and the output end of the fourth feature addition layer is connected to the second input end of the feature subtraction layer;

[0140] The thirteenth RconvBlock module, the fourteenth RconvBlock module and the fifth feature addition layer are connected in sequence, the output end of the fourteenth RconvBlock module is connected to the first input end of the fifth feature addition layer, the output end of the thirteenth RconvBlock module is also connected to the second input end of the sixth feature addition layer, and the output end of the fifth feature addition layer is connected to the first input end of the seventh feature addition layer;

[0141] The fifteenth RconvBlock module, the sixteenth RconvBlock module and the sixth feature addition layer are connected in sequence, the output end of the sixteenth RconvBlock module is connected to the first input end of the sixth feature addition layer, and the output end of the sixth feature addition layer is connected to the second input end of the seventh feature addition layer;

[0142] The output end of the feature subtraction layer is connected to the input end of the twelfth RconvBlock module, and the output end of the twelfth RconvBlock module is connected to the first input end of the eighth feature addition layer; the output end of the seventh feature addition layer is connected to the input end of the seventeenth RconvBlock module, and the output end of the seventeenth RconvBlock module is connected to the second input end of the eighth feature addition layer;

[0143] The output end of the eighth feature addition layer is connected to the input end of the fifth Concat layer, the input end of the sixth Concat layer and the input end of the global average pooling layer respectively, and the output end of the global average pooling layer is connected to the second input end of the first feature multiplication layer and the second input end of the fourth feature multiplication layer respectively; the first feature multiplication layer, the fifth Concat layer and the tenth convolution layer are connected in sequence; the second feature multiplication layer, the sixth Concat layer and the eleventh convolution layer are connected in sequence; the output end of the tenth convolution layer and the output end of the eleventh convolution layer serve as the first and second output ends of the feature fusion unit respectively.

[0144] In this embodiment, the first RConvBlock module, the second RConvBlock module, the third RConvBlock module, the fourth RConvBlock module, the fifth RConvBlock module, the sixth RConvBlock module, the seventh RConvBlock module, the eighth RConvBlock module, the ninth RConvBlock module, the tenth RConvBlock module, the eleventh RConvBlock module, the twelfth RConvBlock module, the thirteenth RConvBlock module, the fourteenth RConvBlock module, the fifteenth RConvBlock module, the sixteenth RConvBlock module and the seventeenth RConvBlock module have the same structure. Figure 8, which is a schematic diagram of the structure of the RConvBlock module, including a fifth DwConv layer, a fifth ReLU activation function, a fifth batch normalization layer, a sixth DwConv layer, a sixth ReLU activation function, a sixth batch normalization layer and a seventh Concat layer connected in sequence;

[0145] The output end of the fifth DwConv layer and the output end of the fifth batch normalization layer are also connected to the input end of the seventh Concat layer; the input end of the fifth DwConv serves as the input end of each of the RConvBlock modules, and the output end of the seventh Concat layer serves as the output end of each of the RConvBlock modules.

[0146] In the feature fusion unit, input 1 and input 2 represent the feature maps of the two branches respectively. The unit is divided into a sum branch and a difference branch, and the feature maps are input into the two branches respectively. Each branch is constructed by two convolution modules with weight sharing, namely RconvBlock modules. After the convolution operation integrates the features, the outputs of the two branches in each branch are respectively added and subtracted element by element, which correspond to the seventh feature addition layer and feature subtraction layer respectively. Among them, the sum branch fuses the features from two different feature maps through the addition operation of the seventh feature addition layer, enhancing the edge information in the image; while the difference branch generates features reflecting the changes through the subtraction operation of the feature subtraction layer, directly revealing the difference between the two feature maps, thereby highlighting the target area and reducing the interference of background noise, further improving the accuracy of target detection.

[0147] Next, the feature fusion output is converted into channel attention coefficients using a global average pooling layer, and the attention operation is performed on the original input image. The output image after the attention operation is concatenated with the feature fusion output to generate a final output image with richer features.

[0148] In the feature fusion unit, the feature subtraction layer output of the difference branch can be expressed as:

[0149] F M =abs(Conv(Conv(L1)+L1)-Conv(Conv(L2)+L2))

[0150] The output of the seventh feature addition layer of the sum branch can be expressed as:

[0151] F P =Conv(Conv(L1)+L1)+Conv(Conv(L2)+L2)

[0152] Among them, abs(·) represents the absolute value operation, and Conv(·) represents the RconvBlock module operation.

[0153] The two outputs of the feature fusion unit can be expressed as:

[0154] Output1=Conv(Concat(GAP(F M +F P )×F1,(F M +F P )))

[0155] Output2 = Conv(Concat(GAP(F M +F P )×F2,(F M +F P )))

[0156] Among them, GAP(·) represents the global average pooling layer, and Concat(·) represents the Concat layer, which concatenates different feature maps along the channel level.

[0157] The neutralization branch of the feature fusion unit fuses features from different feature maps through addition operations, enhancing the edge information of the image; the difference branch generates features that reflect changes through differential operations, revealing the differences between feature maps. This method not only reduces the problem of feature misalignment, but also calculates more accurate features, thereby effectively highlighting the target area, weakening the impact of background noise, and further improving the accuracy of subsequent processing.

[0158] (4) Decoder

[0159] After obtaining the features extracted by the encoder, it is necessary to decode them. Here, a decoder based on the Mamba module is used.

[0160] like Fig. 9 , which is a schematic diagram of the structure of the decoder, wherein the decoder includes a ninth convolutional layer, a first linear projection submodule, a first normalization layer, a first normalization layer, a third Mamba module, a first feature addition layer, a second normalization layer, a first multilayer perceptron and a second feature addition layer connected in sequence;

[0161] The output end of the third Mamba module is connected to the first input end of the first feature addition layer, and the input end of the first multilayer perceptron is connected to the first input end of the second feature addition layer; the output end of the first normalization layer is connected to the second input end of the first feature addition layer, and the output end of the first feature addition layer is also connected to the second input end of the second feature addition layer; the input end of the ninth convolutional layer serves as the input end of the decoder, and the output end of the second feature addition layer serves as the output end of the decoder.

[0162] In the decoder, the feature maps output by the two encoder branches are first concatenated at the channel level. Then, before the feature map is input to the decoder, it needs to be projected similarly to the Vison Transformer, i.e., the linear projection submodule, to project the original image into a vector of fixed length. Assume that the input feature size is X∈R H×W×3 , use 3×3 convolution to project the feature map. Assuming that the size of each patch is n×n, the size of the feature map after projection becomes The projected feature map is then layer normalized and fed into the third Mamba module.

[0163] In this embodiment, the first Mamba module, the second Mamba module and the third Mamba module have the same structure. Fig.10 , which is a schematic diagram of the structure of the Mamba module, including a second linear projection submodule, a third linear projection submodule, a fourth linear projection submodule, a twelfth convolutional layer, a first SiLU activation function, a second SiLU activation function, a state space submodule and a second normalization layer;

[0164] The second linear projection submodule, the twelfth convolutional layer, the first SiLU activation function, the state space submodule, the second normalization layer and the fourth linear projection submodule are connected in sequence;

[0165] The input end of the second linear projection submodule is used as the input end of each of the Mamba modules, the input end of the third linear projection submodule is connected to the input end of the first linear projection submodule, the output end of the third linear projection submodule is connected to the input end of the second SiLU activation function, and the output end of the second SiLU activation function is connected to the input end of the second normalization layer;

[0166] The output end of the fourth linear projection submodule serves as the output end of each of the Mamba modules.

[0167] In the Mamba module, the feature vector is projected again. The features of the left branch are then processed through a convolutional layer and SiLU activation function and input into the state-space model to extract the global context features of the image. Finally, the output vector is dimensionally adjusted through a multi-layer perceptron (MLP) to generate the final decoded output. After being processed by two decoders based on the Mamba architecture, the feature map output by the decoder is passed through two convolutional layers to obtain the final mask detection result.

[0168] By adopting a decoder based on the Mamba module, the network can utilize the global receptive field and dynamic weighting, alleviate the modeling limitations of traditional convolutional neural networks, and provide more advanced modeling capabilities, so that it can effectively detect weak targets in the background of space. The loss function of the weak heat source target detection network model described in this embodiment is:

[0169] L=L det +λ·L enhance

[0170] Among them, λ represents the weight, the first part L det It is the IoU loss function between the segmentation map output by the detection model and the true segmentation map:

[0171]

[0172] Among them, y represents the segmentation map output by the detection model, x gt Represents the true segmentation map, and λ represents a pre-set constant used to balance the ratio of the two parts of the loss function.

[0173] In the specific implementation process, this embodiment first designs an image enhancement unit, first uses a CNN-based multi-scale feature extraction module and Mamba architecture to extract local and global features of the input image in parallel. By combining local and global features, the module can understand the image content more comprehensively, thereby retaining important details during the enhancement process and avoiding artifacts or distortion caused by over-enhancement. Subsequently, a CNN containing a jump connection is used for image restoration. The use of jump connections during the restoration process can retain small target details as much as possible, and obtain an enhanced image with enhanced contrast and brightness, which helps to highlight the important information of the target in the image. The architecture is based on a U-shaped network structure and performs feature extraction at three different levels to capture the multi-scale features of the image.

[0174] After being enhanced by the image enhancement unit, the original image and the enhanced image are simultaneously input into the dual-branch parallel encoder for feature extraction. In this process, by setting feature jump connections at different levels and depths, it can help restore the details in the image, especially to prevent the details of small targets from being lost during multiple convolutions and downsampling. The enhanced image usually has better contrast and brightness, making the target more prominent in the image, while the original image may retain more detail information. Through dual-path parallel feature extraction, the detail features of the original image and the enhanced image can be comprehensively obtained, thereby improving the quality of the overall features and ultimately improving the detection performance of weak targets. Subsequently, the feature fusion unit is used to integrate the feature information of the original input image and the enhanced image. Since these two types of images contain different details, the original image retains richer original details, while the enhanced image has higher contrast and brightness after preprocessing. Fusion of the features of these two types of images can make full use of their respective advantages and improve the comprehensiveness and expression ability of the overall features. The output result after feature fusion is input into the feature embedding unit, and then decoded using a decoder based on the Mamba architecture. This step is crucial for the detection of infrared weak targets, especially in an environment with complex space background and many interference factors. The decoder of the Mamba architecture can efficiently process the multi-scale features extracted by the encoder, ensuring that the integrity of key details is maintained during the decoding process. A significant advantage of the Mamba architecture is its multi-scale feature fusion capability. During the decoding process, it is able to combine details and global information at different scales to ensure that the model does not lose the position and context of the target in the overall image when refining small targets. This multi-scale fusion strategy helps improve the model's ability to recognize and locate targets at different scales, especially when processing complex infrared images.

[0175] In addition, infrared images are often affected by noise, especially in space environments. The decoder based on the Mamba architecture effectively suppresses the influence of noise through fine processing and feature enhancement. It can improve the contrast of the target, making weak infrared heat source targets more clearly visible in the image. This feature enhancement and noise suppression mechanism not only improves the visibility of the image, but also provides clearer and more reliable input for accurate detection by subsequent detection units. Ultimately, the decoder based on the Mamba architecture can ensure that the model outputs clear and accurate detection results. In the complex space environment, it can effectively cope with various noises and background interference, provide strong support for the detection of weak infrared targets, and ensure that the detection unit can generate accurate target positioning and classification results.

[0176] Example 3

[0177] This embodiment provides a weak heat source target detection system under a space background, which is used to implement the weak heat source target detection method under a space background described in Embodiment 1 or Embodiment 2, such as Fig.11 As shown, including:

[0178] A data acquisition module is used to acquire several infrared images of space background with real heat source annotations;

[0179] The first detection module is used to input the space background infrared image with real heat source annotations into the constructed weak heat source target detection network model to obtain the heat source detection result in the space background infrared image;

[0180] A model training module is used to establish a total loss function based on the heat source detection results and the real heat source annotations, optimize the constructed weak heat source target detection network model, and obtain a trained weak heat source target detection network model when the total loss function value reaches the minimum;

[0181] The second detection module is used to obtain the space background infrared image to be detected, input it into the trained weak heat source target detection network model, and obtain the heat source detection result in the space background infrared image to be detected.

[0182] The same or similar reference numerals correspond to the same or similar components;

[0183] The terms used in the drawings to describe positional relationships are only used for illustrative purposes and should not be construed as limiting this patent;

[0184] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A method for detecting weak heat source targets in space background, characterized in that: include: Obtain several space background infrared images with real heat source annotations; Inputting the space background infrared image with real heat source annotations into the constructed weak heat source target detection network model to obtain the heat source detection result in the space background infrared image; A total loss function is established based on the heat source detection results and the real heat source annotations, and the constructed weak heat source target detection network model is optimized. When the total loss function value reaches the minimum, a trained weak heat source target detection network model is obtained; Obtain the space background infrared image to be detected, input it into the trained weak heat source target detection network model, and obtain the heat source detection result in the space background infrared image to be detected.

2. The method for detecting weak heat source targets under space background according to claim 2 is characterized in that: The weak heat source target detection network model includes an image enhancement unit, a first encoder, a second encoder, a third encoder, a fourth encoder, a fifth encoder, a sixth encoder, a first feature fusion unit, a second feature fusion unit, a third feature fusion unit, a feature embedding unit, a first decoder, a second decoder, a detection unit, a first feature splicing layer, a second feature splicing layer, a third feature splicing layer, a fourth feature splicing layer and a fifth feature splicing layer; The input end of the image enhancement unit is connected to the input end of the first encoder, the output end of the image enhancement unit is connected to the input end of the second encoder, the output end of the second encoder is respectively connected to the second input end of the first feature fusion unit and the input end of the second feature splicing layer, the output end of the first encoder is respectively linked to the first input end of the first feature fusion unit and the input end of the first feature splicing layer, the first and second output ends of the first feature fusion unit are respectively connected to the input end of the first feature splicing layer and the input end of the second feature splicing layer, the output end of the first feature splicing layer is connected to the input end of the third encoder, and the output end of the second feature splicing layer is connected to the input end of the fourth encoder; The output end of the third encoder is connected to the first input end of the second feature fusion unit and the input end of the third feature splicing layer respectively, the output end of the fourth encoder is connected to the second input end of the second feature fusion unit and the input end of the fourth feature splicing layer respectively, the first and second output ends of the second feature fusion unit are connected to the input end of the third feature splicing layer and the input end of the fourth feature splicing layer respectively, the output end of the third feature splicing layer is connected to the input end of the fifth encoder, and the output end of the fourth feature splicing layer is connected to the input end of the sixth encoder; The output end of the fifth encoder is connected to the first input end of the third feature fusion unit, the output end of the sixth encoder is connected to the second input end of the second feature fusion unit, the first and second output ends of the third feature fusion unit are respectively connected to the input end of the fifth feature splicing layer, the output end of the fifth feature splicing layer is connected to the input end of the feature embedding unit, and the feature embedding unit, the first decoder, the second decoder and the detection unit are connected in sequence.

3. The method for detecting weak heat source targets under space background according to claim 2, characterized in that: The image enhancement unit includes a first Mamba module, a second Mamba module, and a first MconvBlock module, a second MconvBlock module, a third MconvBlock module, a first Concat layer, a first DWconv layer, a first RConvBlock module, a second RConvBlock module, a third RConvBlock module and a first FConvBlock module connected in sequence; The input end of the first MconvBlock module is used as the input end of the image enhancement unit, the input end of the first Mamba module is connected to the input end of the first MconvBlock module, the output end of the first Mamba module is connected to the input end of the second Mamba module, the output end of the second Mamba module is connected to the input end of the first Concat layer, and the output ends of the first RConvBlock module and the second RConvBlock module are also connected to the input end of the first FConvBlock module; The output end of the first FConvBlock module serves as the output end of the image enhancement unit.

4. The method for detecting weak heat source targets under space background according to claim 3, characterized in that: The first MconvBlock module, the second MconvBlock module, and the third MconvBlock module have the same structure, including a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a fifth convolutional layer, a sixth convolutional layer, a seventh convolutional layer, a first average pooling layer, a sixth feature concatenation layer, a first batch of normalization layers, and a first ReLU activation function; The input end of the first convolutional layer is used as the input end of the MconvBlock module, the output end of the first convolutional layer is connected to the input ends of the second convolutional layer, the third convolutional layer, the fourth convolutional layer and the first average pooling layer, the output end of the third convolutional layer is connected to the input end of the fifth convolutional layer, the output end of the fourth convolutional layer is connected to the input end of the sixth convolutional layer, the input end of the first average pooling layer is connected to the input end of the seventh convolutional layer, and the output ends of the second convolutional layer, the fifth convolutional layer, the sixth convolutional layer and the seventh convolutional layer are all connected to the input end of the sixth feature splicing layer; The sixth feature concatenation layer, the first normalization layer, and the first ReLU activation function are connected in sequence; The output of the first ReLU activation function is used as the output of the MconvBlock module.

5. The method for detecting weak heat source targets under space background according to claim 3, characterized in that: The FConvBlock module includes a second Concat layer, a second DwConv layer, a second ReLU activation function, a second batch normalization layer, a third DwConv layer, a third ReLU activation function, a third batch normalization layer and a third Concat layer connected in sequence; The output of the second DwConv layer and the output of the second batch normalization layer are also connected to the input of the third Concat layer; The input end of the second Concat layer is used as the input end of the FConvBlock module, and the output end of the third Concat layer is used as the output end of the FConvBlock module.

6. The method for detecting weak heat source targets under space background according to claim 2, characterized in that: The encoder includes a fourth RconvBlock module, a fifth RconvBlock module, a sixth RconvBlock module, a seventh RconvBlock module, a first AConvBlock module and a second AConvBlock module; The fourth RconvBlock module, the fifth RconvBlock module and the first AConvBlock module are connected in sequence, and the output end of the fourth RconvBlock module is also connected to the input end of the first AConvBlock module and the input end of the sixth RconvBlock module; The output end of the sixth RconvBlock module is connected to the input end of the seventh RconvBlock module and the input end of the second AConvBlock module respectively, the output end of the seventh RconvBlock module is connected to the input end of the second AConvBlock module, and the output end of the second AConvBlock module is connected to the input end of the first AConvBlock module; The input end of the fourth RconvBlock module is used as the input end of the encoder, and the output end of the first AConvBlock module is used as the output end of the encoder; The AConvBlock module includes a fourth Concat layer, an eighth convolutional layer, a channel space attention submodule, a fourth DwConv layer, a fourth ReLU activation function and a fourth batch normalization layer connected in sequence; The input end of the fourth Concat layer is used as the input end of the AConvBlock module, and the output end of the fourth batch normalization layer is used as the output end of the AConvBlock module.

7. The method for detecting weak heat source targets under space background according to claim 2, characterized in that: The decoder comprises a ninth convolutional layer, a first linear projection submodule, a first normalization layer, a first normalization layer, a third Mamba module, a first feature addition layer, a second normalization layer, a first multilayer perceptron and a second feature addition layer connected in sequence; The output end of the third Mamba module is connected to the first input end of the first feature addition layer, and the input end of the first multilayer perceptron is connected to the first input end of the second feature addition layer; The output end of the first normalization layer is connected to the second input end of the first feature addition layer, and the output end of the first feature addition layer is also connected to the second input end of the second feature addition layer; The input end of the ninth convolutional layer is used as the input end of the decoder, and the output end of the second feature addition layer is used as the output end of the decoder.

8. The method for detecting weak heat source targets under space background according to claim 2, characterized in that: The feature fusion unit includes an eighth RconvBlock module, a ninth RconvBlock module, a tenth RconvBlock module, an eleventh RconvBlock module, a twelfth RconvBlock module, a thirteenth RconvBlock module, a fourteenth RconvBlock module, a fifteenth RconvBlock module, a sixteenth RconvBlock module, a seventeenth RconvBlock module, a third feature addition layer, a fourth feature addition layer, a fifth feature addition layer, a sixth feature addition layer, a seventh feature addition layer, an eighth feature addition layer, a first feature multiplication layer, a second feature multiplication layer, a feature subtraction layer, a global average pooling layer, a fifth Concat layer, a sixth Concat layer, a tenth convolution layer and an eleventh convolution layer; The input end of the eighth RconvBlock module is used as the first input end of the feature fusion unit, and the input end of the eighth RconvBlock module is connected to the first input end of the first feature multiplication layer and the input end of the thirteenth RconvBlock module respectively; The input end of the fifteenth RconvBlock module serves as the second input end of the feature fusion unit, and the input end of the fifteenth RconvBlock module is respectively connected to the first input end of the second feature multiplication layer and the input end of the tenth RconvBlock module; The eighth RconvBlock module, the ninth RconvBlock module and the third feature addition layer are connected in sequence, the output end of the ninth RconvBlock module is connected to the first input end of the third feature addition layer, the output end of the eighth RconvBlock module is also connected to the second input end of the third feature addition layer, and the output end of the third feature addition layer is connected to the first input end of the feature subtraction layer; The tenth RconvBlock module, the eleventh RconvBlock module and the fourth feature addition layer are connected in sequence, the output end of the eleventh RconvBlock module is connected to the first input end of the fourth feature addition layer, the output end of the tenth RconvBlock module is also connected to the second input end of the fourth feature addition layer and the second input end of the fifth feature addition layer, and the output end of the fourth feature addition layer is connected to the second input end of the feature subtraction layer; The thirteenth RconvBlock module, the fourteenth RconvBlock module and the fifth feature addition layer are connected in sequence, the output end of the fourteenth RconvBlock module is connected to the first input end of the fifth feature addition layer, the output end of the thirteenth RconvBlock module is also connected to the second input end of the sixth feature addition layer, and the output end of the fifth feature addition layer is connected to the first input end of the seventh feature addition layer; The fifteenth RconvBlock module, the sixteenth RconvBlock module and the sixth feature addition layer are connected in sequence, the output end of the sixteenth RconvBlock module is connected to the first input end of the sixth feature addition layer, and the output end of the sixth feature addition layer is connected to the second input end of the seventh feature addition layer; The output end of the feature subtraction layer is connected to the input end of the twelfth RconvBlock module, and the output end of the twelfth RconvBlock module is connected to the first input end of the eighth feature addition layer; The output end of the seventh feature addition layer is connected to the input end of the seventeenth RconvBlock module, and the output end of the seventeenth RconvBlock module is connected to the second input end of the eighth feature addition layer; The output end of the eighth feature addition layer is connected to the input end of the fifth Concat layer, the input end of the sixth Concat layer and the input end of the global average pooling layer respectively, and the output end of the global average pooling layer is connected to the second input end of the first feature multiplication layer and the second input end of the fourth two-feature multiplication layer respectively; The first feature multiplication layer, the fifth Concat layer, and the tenth convolution layer are connected in sequence; The second feature multiplication layer, the sixth Concat layer, and the eleventh convolution layer are connected in sequence; The output end of the tenth convolutional layer and the output end of the eleventh convolutional layer serve as the first and second output ends of the feature fusion unit, respectively.

9. The method for detecting weak heat source targets under space background according to claim 3, 6 or 8, characterized in that: The first RConvBlock module, the second RConvBlock module, the third RConvBlock module, the fourth RConvBlock module, the fifth RConvBlock module, the sixth RConvBlock module, the seventh RConvBlock module, the eighth RConvBlock module, the ninth RConvBlock module, the tenth RConvBlock module, the eleventh RConvBlock module, the twelfth RConvBlock module, the thirteenth RConvBlock module, the fourteenth RConvBlock module, the fifteenth RConvBlock module, the sixteenth RConvBlock module and the seventeenth RConvBlock module have the same structure, including a fifth DwConv layer, a fifth ReLU activation function, a fifth batch normalization layer, a sixth DwConv layer, a sixth ReLU activation function, a sixth batch normalization layer and a seventh Concat layer connected in sequence; The output of the fifth DwConv layer and the output of the fifth batch normalization layer are also connected to the input of the seventh Concat layer; The input end of the fifth DwConv serves as the input end of each of the RConvBlock modules, and the output end of the seventh Concat layer serves as the output end of each of the RConvBlock modules.

10. The method for detecting weak heat source targets in space background according to claim 3 or 7, characterized in that: The first Mamba module, the second Mamba module and the third Mamba module have the same structure, including a second linear projection submodule, a third linear projection submodule, a fourth linear projection submodule, a twelfth convolutional layer, a first SiLU activation function, a second SiLU activation function, a state space submodule and a second normalization layer; The second linear projection submodule, the twelfth convolutional layer, the first SiLU activation function, the state space submodule, the second normalization layer and the fourth linear projection submodule are connected in sequence; The input end of the second linear projection submodule is used as the input end of each of the Mamba modules, the input end of the third linear projection submodule is connected to the input end of the first linear projection submodule, the output end of the third linear projection submodule is connected to the input end of the second SiLU activation function, and the output end of the second SiLU activation function is connected to the input end of the second normalization layer; The output end of the fourth linear projection submodule serves as the output end of each of the Mamba modules.

Images