Infrared small target detection method and device based on diffusion model data enhancement
By adopting a data augmentation method based on diffusion model in infrared small object detection, the detection problem of infrared small object in complex backgrounds and noise is solved, and the detection efficiency and sample diversity and authenticity are improved.
Patent Information
- Application Number
- CN202411831399.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-12
- Publication Date
- 2025-04-18
AI Technical Summary
The existing infrared small-object detection technology faces the problems of small targets being difficult to detect in complex backgrounds and noise, weak target characteristics, great influence of light changes and environmental factors, resulting in low detection efficiency.
Using a data augmentation method based on the diffusion model, the image to be detected is converted into a stitched image through Mosaic data augmentation, and preprocessed it, input it into the preset image augmentation network and diffusion model, and generate coordinated images and resampled images, and finally perform infrared small object detection through the detection network.
The efficiency of infrared small target detection is improved, and the generated augmented samples have higher diversity and authenticity, and can detect infrared small targets more accurately.
Smart Images
Figure CN119850527A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of target detection, and in particular to a method and device for detecting small infrared targets based on diffusion model data enhancement. Background Art
[0002] In recent years, as accurate detection of small infrared targets has become increasingly important for safety, navigation, and mission execution, the infrared small target detection task has also received widespread attention, especially in the fields of maritime surveillance, early warning systems, precision guidance, and traffic monitoring. However, infrared small target detection faces multiple challenges. First, since small targets occupy very few pixels in infrared images, they are easily submerged in complex backgrounds and noise, making them difficult to detect. Second, small infrared targets usually have no obvious shape and texture, which makes it more difficult to detect them. In addition, the features of these small targets in the image are weak, and accurate detection requires overcoming the influence of illumination changes and other environmental factors. Therefore, how to achieve accurate detection of small infrared targets in the face of these problems remains a challenging problem.
[0003] In order to deal with this problem, many methods for improving the target detection model have been adopted in the existing technology to improve the detection efficiency of infrared small targets. However, researchers are usually limited to using some mainstream data enhancement techniques, such as mosaic, cut-mix and mixup, which generate new images by mixing information from different real images.
[0004] However, due to the small number of samples in the infrared dataset and the small sample gap, these mainstream data enhancement methods are prone to lack of diversity when generating samples. In addition, due to the large difference in contrast between infrared images, the augmented samples generated by mosaic and other methods will appear obviously inconsistent at the boundaries and lack a sense of reality. In turn, this has a significant negative impact on the subsequent infrared small target detection. Summary of the invention
[0005] In order to overcome the above-mentioned defects of the prior art, the present invention provides an infrared small target detection method and device based on diffusion model data enhancement, which can improve the efficiency of infrared small target detection.
[0006] An embodiment of the present invention provides an infrared small target detection method based on diffusion model data enhancement, comprising the following steps: The image to be detected is converted into a spliced image by Mosaic data enhancement, and the spliced image is preprocessed; The preprocessed stitched image is input into a preset image enhancement network to obtain a coordinated image; Inputting the coordinated image into a preset diffusion model to obtain a resampled image; The resampled image is subjected to infrared small target detection through a preset detection network to obtain a detection image.
[0007] Furthermore, the preprocessing of the stitched image specifically includes: Performing degradation processing on the spliced image to obtain a degraded image; A region whose area is a preset area threshold is selected from the stitched image as a clipping region, and the image in the clipping region is cut and pasted to a corresponding position of the degraded image to complete the preprocessing.
[0008] Furthermore, the preprocessed stitched image is input into a preset image enhancement network to obtain a coordinated image, which specifically includes: The preset image enhancement network includes a shallow convolution module, a residual module and a reconstruction module; After the preprocessed spliced image is input into the preset image enhancement network, shallow features are extracted by the shallow convolution module, and deep features are extracted by the residual module; The shallow features and the deep features are fused through the reconstruction module, and the fused features are upsampled to obtain the coordinated image.
[0009] Preferably, before the preprocessed spliced image is input into the preset image enhancement network, the method further includes: optimizing the preset image enhancement network by using a preset optimized image and a preset enhancement loss function; wherein the optimization process is specifically as follows: Preprocessing the preset optimized image, and inputting the preprocessed preset optimized image into the preset image enhancement network to obtain an optimized coordinated image; According to the optimized coordinated image and the preset optimized image, the preset image enhancement network is optimized by the preset enhancement loss function; wherein the preset enhancement loss function is specifically:
[0010] in, For said optimized coordinated image, Optimizes the image for the preset in question.
[0011] Furthermore, the step of inputting the coordinated image into a preset diffusion model to obtain a resampled image specifically includes: The preset diffusion model includes an encoder, a noise adding module, a denoising network and a decoder; After the coordinated image is input into the preset diffusion model, the coordinated image is encoded into a latent space code by an encoder; Adding noise to the latent space code by the noise adding module to obtain diffuse noise; Resampling the diffuse noise through the denoising network to obtain denoising code; The denoising code is decoded by the decoder to obtain the resampled image.
[0012] Preferably, before the diffuse noise is resampled by the denoising network, the method further includes: training the preset denoising network by using a preset training image and a preset denoising loss function; wherein the training process is specifically as follows: Encoding the preset training image into a training code by the encoder; Adding noise to the training code by the noise adding module to obtain training noise; According to the training noise, the preset denoising network is trained by the preset denoising loss function; wherein the preset denoising loss function is specifically:
[0013] in, is Gaussian noise, Encode the training, is the training noise, For the preset denoising network, is the preset time step, Enter the preset conditions. represents a random variable with uniform distribution U(0,1) and t Take the expected value.
[0014] Furthermore, the step of performing infrared small target detection on the resampled image through a preset detection network to obtain a detection image specifically includes: The preset detection network includes several residual attention blocks and feature fusion modules; After the resampled image is input into the preset detection network, features of the resampled image are extracted through the plurality of residual attention blocks to obtain a feature pyramid; The feature fusion module performs feature fusion on the feature pyramid to obtain the detection image.
[0015] Preferably, after obtaining the detection image, the method further includes: optimizing the preset detection network according to the detection image and a preset detection loss function; wherein the preset detection loss function is specifically:
[0016] in, is the detection image, To presume the real result, is the preset smoothing coefficient.
[0017] Another embodiment of the present invention provides an infrared small target detection device based on diffusion model data enhancement, including: a splicing module, an enhancement module, a diffusion module and a detection module; The stitching module is used to convert the image to be detected into a stitched image through Mosaic data enhancement, and preprocess the stitched image; The enhancement module is used to input the preprocessed spliced image into a preset image enhancement network to obtain a coordinated image; The diffusion module is used to input the coordinated image into a preset diffusion model to obtain a resampled image; The detection module is used to perform infrared small target detection on the resampled image through a preset detection network to obtain a detection image.
[0018] Furthermore, the enhancement module is used to input the preprocessed stitched image into a preset image enhancement network to obtain a coordinated image, which specifically includes: The preset image enhancement network includes a shallow convolution module, a residual module and a reconstruction module; After the preprocessed spliced image is input into the preset image enhancement network, shallow features are extracted by the shallow convolution module, and deep features are extracted by the residual module; The shallow features and the deep features are fused through the reconstruction module, and the fused features are upsampled to obtain the coordinated image.
[0019] Compared with the prior art, the present invention has the following beneficial effects: Based on the improvement of traditional mosaic data enhancement, an image enhancement network based on pixel space is proposed. The image enhancement network can coordinate images without labels to generate high-quality and realistic mosaic images. At the same time, the present invention also proposes a diffusion model based on diffusion prior. The diffusion prior method is introduced on the infrared data set, and the pixel prior generation results are resampled by diffusion generation prior, and the knowledge of the real world is integrated. This makes the samples generated by the diffusion prior more realistic, and because of the integration of real-world knowledge, the generated samples are more diverse.
[0020] In summary, the present invention improves the diversity and authenticity of augmented samples by improving the traditional data enhancement method, thereby improving the detection efficiency of small infrared targets. BRIEF DESCRIPTION OF THE DRAWINGS
[0021] Figure 1A flowchart of an infrared small target detection method based on diffusion model data enhancement is provided in accordance with an embodiment of the present invention.
[0022] Figure 2 A schematic diagram of an optimization process of a preset image enhancement network provided by an embodiment of the present invention.
[0023] Figure 3 A schematic diagram of an optimization process of a preset diffusion model provided by an embodiment of the present invention.
[0024] Figure 4 A flowchart of optimizing training of a preset detection network provided by one embodiment of the present invention.
[0025] Figure 5 A schematic diagram of test results of an infrared small target detection method based on diffusion model data enhancement on the SIRST dataset provided by an embodiment of the present invention.
[0026] Figure 6 A schematic diagram of test results of an infrared small target detection method based on diffusion model data enhancement on the NUDT-SIRST dataset provided by an embodiment of the present invention.
[0027] Figure 7 A schematic structural diagram of an infrared small target detection device based on diffusion model data enhancement provided by another embodiment of the present invention. DETAILED DESCRIPTION
[0028] The drawings are for illustrative purposes only and should not be construed as limiting the present patent; It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.
[0029] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.
[0030] Reference Figure 1 , is a flow chart of a method for detecting small infrared targets based on diffusion model data enhancement provided by an embodiment of the present invention, comprising the following steps: S1: converting the image to be detected into a spliced image through Mosaic data enhancement, and preprocessing the spliced image; S2: input the preprocessed stitched image into a preset image enhancement network to obtain a coordinated image; S3: inputting the coordinated image into a preset diffusion model to obtain a resampled image; S4: Perform infrared small target detection on the resampled image through a preset detection network to obtain a detection image.
[0031] For step S1, specifically, the preprocessing of the stitched image specifically includes: Performing degradation processing on the spliced image to obtain a degraded image; A region whose area is a preset area threshold is selected from the stitched image as a clipping region, and the image in the clipping region is cut and pasted to a corresponding position of the degraded image to complete the preprocessing.
[0032] In a preferred embodiment, the images to be detected are first spliced into the spliced image using a traditional mosaic strategy. Then, in order to improve the image quality of the image restoration process, the preferred embodiment uses a cut-and-paste operation. The cut-and-paste operation can increase the diversity and difficulty of the data, so that the image quality of the restored model is higher.
[0033] Before performing the clipping operation, the stitched image needs to be degraded. The degradation process is as follows:
[0034] in, is the degraded image, It represents the degradation module, which uses a variety of degradation operators, including blur, resizing, noise and other degradation methods. is the stitched image.
[0035] After obtaining the degraded image, a part of the stitched image is pasted into the degraded image, which is specifically expressed as follows:
[0036] in, (·) indicates a cut and paste operation. Represents the selected clipping area from the stitched image.
[0037] For step S2, specifically, the step of inputting the preprocessed stitched image into a preset image enhancement network to obtain a coordinated image specifically includes: The preset image enhancement network includes a shallow convolution module, a residual module and a reconstruction module; After the preprocessed spliced image is input into the preset image enhancement network, shallow features are extracted by the shallow convolution module, and deep features are extracted by the residual module; The shallow features and the deep features are fused through the reconstruction module, and the fused features are upsampled to obtain the coordinated image.
[0038] Preferably, before the preprocessed spliced image is input into the preset image enhancement network, the method further includes: optimizing the preset image enhancement network by using a preset optimized image and a preset enhancement loss function; wherein the optimization process is specifically as follows: Preprocessing the preset optimized image, and inputting the preprocessed preset optimized image into the preset image enhancement network to obtain an optimized coordinated image; According to the optimized coordinated image and the preset optimized image, the preset image enhancement network is optimized by the preset enhancement loss function; wherein the preset enhancement loss function is specifically:
[0039] in, For said optimized coordinated image, Optimizes the image for the preset in question.
[0040] In a preferred embodiment, after obtaining the preprocessed stitched image, the preprocessed stitched image needs to be input into a transformer network (ie, the preset image enhancement network) for restoration. Before formal restoration, the transformer network needs to be optimized.
[0041] Reference Figure 2 , is a schematic diagram of an optimization process of a preset image enhancement network provided by an embodiment of the present invention. Figure 2 It can be seen that, first of all, the preset optimization image needs to be The preprocessing process is the same as that of the stitched image. Degrade and get the degraded image Then, the mixed image is obtained by clipping and pasting Then The image is input into the preset image enhancement network, and shallow features are extracted through the shallow convolution module, and deep features are extracted through the residual module; wherein the residual module includes multiple Swin transformer layers, which work together to capture the deeper features of the image. Then, we fuse the shallow and deep feature maps to integrate high-frequency and low-frequency information. Finally, the fused features are restored to the original image space through upsampling operation to obtain the coordinated image. In order to optimize the network, it is necessary to calculate the preset enhanced loss function and The preset image enhancement network is optimized with the goal of minimizing the gap.
[0042] In the actual working stage, the preprocessed stitched image is input into the optimized preset image enhancement network to output the coordinated image.
[0043] Regarding step S3, specifically, inputting the coordinated image into a preset diffusion model to obtain a resampled image specifically includes: The preset diffusion model includes an encoder, a noise adding module, a denoising network and a decoder; After the coordinated image is input into the preset diffusion model, the coordinated image is encoded into a latent space code by an encoder; Adding noise to the latent space code by the noise adding module to obtain diffuse noise; Resampling the diffuse noise through the denoising network to obtain denoising code; The denoising code is decoded by the decoder to obtain the resampled image.
[0044] In a preferred embodiment, the coordinated image obtained by pixel prior is smoother and more coordinated than the spliced distribution, but there is still a gap between the coordinated image and the real image. This is because the enhancement network is optimized according to the loss function, and the use of L2 loss will amplify the gap of small changes in feature values. Therefore, the coordinated image obtained by pixel prior pays more attention to the consistency of the overall pixel value, while ignoring the details in the real image.
[0045] In order to solve this problem, the present preferred embodiment introduces the prior of the diffusion model. Different from the traditional data enhancement method, the generated fictitious data lacks realism and is difficult to capture the complex structure and potential characteristics of the original data. Using the diffusion model prior as a prior fully considers the inherent structure of the original data, so that the data distribution can be simulated more effectively. In the present preferred embodiment, the potential diffusion model is used as the preset diffusion model to generate the prior.
[0046] After obtaining the coordinated image, the coordinated image needs to be input into the preset diffusion model for resampling. Before formal resampling, the preset diffusion model needs to be optimized.
[0047] Reference Figure 3 , is a schematic diagram of an optimization process of a preset diffusion model provided by an embodiment of the present invention. First, the image Encoding into latent space encoding Then, the noise module is used to add Introduce T-step noise to make it close to Gaussian distribution , and then generate by T-step resampling through the denoising network , and finally the decoder D resamples to generate the image .
[0048] In order to make the diffusion model resampling results more realistic, it is also necessary to To reduce and The gap between them, the loss is expressed as follows:
[0049] After large-scale training, the diffusion model will understand the shape, texture and other features of the objects in the image and generate visually pleasing images. Therefore, the diffusion model is introduced as a priori to model the shape and lighting information of the objects in the image in the real world, and the introduction of real-world information helps generate detailed and realistic images.
[0050] Therefore, in the actual working stage, the coordinated image is input into the optimized preset diffusion model, and the resampled image can be output.
[0051] Preferably, before the diffuse noise is resampled by the denoising network, the method further includes: training the preset denoising network by using a preset training image and a preset denoising loss function; wherein the training process is specifically as follows: Encoding the preset training image into a training code by the encoder; Adding noise to the training code by the noise adding module to obtain training noise; According to the training noise, the preset denoising network is trained by the preset denoising loss function; wherein the preset denoising loss function is specifically:
[0052] in, is Gaussian noise, Encode the training, is the training noise, For the preset denoising network, is the preset time step, Enter the preset conditions. represents a random variable with uniform distribution U(0,1) and t Take the expected value.
[0053] In a preferred embodiment, the purpose of training the denoising network before the actual working phase is to minimize the difference between the predicted noise and the real noise during the inversion process.
[0054] For step S4, specifically, performing infrared small target detection on the resampled image through a preset detection network to obtain a detection image specifically includes: The preset detection network includes several residual attention blocks and feature fusion modules; After the resampled image is input into the preset detection network, features of the resampled image are extracted through the plurality of residual attention blocks to obtain a feature pyramid; The feature fusion module performs feature fusion on the feature pyramid to obtain the detection image.
[0055] Preferably, after obtaining the detection image, the method further includes: optimizing the preset detection network according to the detection image and a preset detection loss function; wherein the preset detection loss function is specifically:
[0056] in, is the detection image, To presume the real result, is the preset smoothing coefficient.
[0057] In a preferred embodiment, referring to Figure 4 , is a flow chart of optimization training of a preset detection network provided by an embodiment of the present invention. Figure 4 It can be seen that when the resampled image is obtained After that, Input to the residual attention block to extract features. In order to fuse different features, the features are continuously upsampled and downsampled and input to different residual attention blocks to form a feature pyramid. Finally, these features containing information of different scales are fused together to generate a robust feature map for prediction results. . Actual target results and prediction results They are all binary images. In order to obtain results that match the real target As a result, the network passes the loss To reduce With real results The gap between.
[0058] In order to prove that the method of the present invention can significantly improve the performance of the detection network, an example is given below using a specific infrared data set and infrared detection network parameters.
[0059] Assume that the parameters of the infrared dataset and the infrared detection network are as follows: The infrared dataset includes the SIRST dataset and the NUDT-SIRST dataset. SIRST is a small open single-frame infrared small target detection dataset that selects images from sequences. The dataset was designated as a public dataset by the University of Arizona in 2020. It contains representative images of 427 different scenes from real-world videos of hundreds of different scenes. These images were taken at shortwave, medium wavelength, and 950nm wavelengths. They are annotated in five different forms to support models for detection tasks and segmentation tasks. The NUDT-SIRST dataset is an artificially synthesized dataset that includes 5 main background scenes, including cities, fields, highlights, seas, and clouds. Each image is synthesized from a real background, with different targets (such as points, points, and extended targets), different SCRs, and rich postures. The total amount of data is 1362.
[0060] The testing process is as follows: 1. Dataset division: Since the NUDT dataset is large, it can be used to prove the effectiveness and robustness of the present invention. We divide the dataset into 50% for training and 50% for testing. SIRST is a small sample dataset that can be used to verify the advantages of the present invention on small sample datasets. We use 70% of the SIRST dataset for training and 30% for testing.
[0061] 2. Network training stage: The present invention uses second-order degradation in the degradation part of the pixel prior. Repeat the blur-resizing-noise process twice for the image input. For the restoration network, the present invention sets the batch size of the transformer network to 4, the training cycle to 100, and the learning rate to 0.001. The present invention uses a diffusion model pre-trained on ImageNet, and sets the batch size, training cycle, and learning rate to 2, 100, and 0.01, respectively, for fine-tuning. The present invention generates 398 samples on the NUDT-SIRST dataset and 100 samples on the SIRST dataset. These samples are input into the baseline network as additional training samples. The baseline detection network is trained with a batch size of 40 and a training cycle of 3000.
[0062] 3. Test set test phase: In order to evaluate the performance of the present invention, the present invention uses IoU, detection probability ( ) and false positive rate ( ) as the evaluation indicator. The trained detection model obtained in the above steps is tested on the NUDT and SIRST test sets. Figure 5 and Figure 6The test results are respectively on the SIRST dataset and the NUDT-SIRST dataset. As can be seen from the figure, compared with other currently more advanced infrared small target detection methods. During the test, the detection model after training the detection network using the data generated by the present invention performed well in predicting infrared small targets, and there would be no missed detection or false detection.
[0063] Reference Figure 7 , is a schematic structural diagram of an infrared small target detection device based on diffusion model data enhancement provided by another embodiment of the present invention, comprising: a splicing module 101, an enhancement module 102, a diffusion module 103 and a detection module 104; The stitching module 101 is used to convert the image to be detected into a stitched image through Mosaic data enhancement, and pre-process the stitched image; The enhancement module 102 is used to input the preprocessed spliced image into a preset image enhancement network to obtain a coordinated image; The diffusion module 103 is used to input the coordinated image into a preset diffusion model to obtain a resampled image; The detection module 104 is used to perform infrared small target detection on the resampled image through a preset detection network to obtain a detection image.
[0064] Furthermore, the enhancement module 102 is used to input the preprocessed stitched image into a preset image enhancement network to obtain a coordinated image, which specifically includes: The preset image enhancement network includes a shallow convolution module, a residual module and a reconstruction module; After the preprocessed spliced image is input into the preset image enhancement network, shallow features are extracted by the shallow convolution module, and deep features are extracted by the residual module; The shallow features and the deep features are fused through the reconstruction module, and the fused features are upsampled to obtain the coordinated image.
[0065] 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 small infrared targets based on diffusion model data enhancement, characterized in that: The steps include: The image to be detected is converted into a spliced image by Mosaic data enhancement, and the spliced image is preprocessed; The preprocessed stitched image is input into a preset image enhancement network to obtain a coordinated image; Inputting the coordinated image into a preset diffusion model to obtain a resampled image; The resampled image is subjected to infrared small target detection through a preset detection network to obtain a detection image.
2. The infrared small target detection method based on diffusion model data enhancement as claimed in claim 1, characterized in that: The preprocessing of the stitched image specifically includes: Performing degradation processing on the spliced image to obtain a degraded image; A region whose area is a preset area threshold is selected from the stitched image as a clipping region, and the image in the clipping region is cut and pasted to a corresponding position of the degraded image to complete the preprocessing.
3. The infrared small target detection method based on diffusion model data enhancement as claimed in claim 1, characterized in that: The preprocessed stitched image is input into a preset image enhancement network to obtain a coordinated image, specifically comprising: The preset image enhancement network includes a shallow convolution module, a residual module and a reconstruction module; After the preprocessed spliced image is input into the preset image enhancement network, shallow features are extracted by the shallow convolution module, and deep features are extracted by the residual module; The shallow features and the deep features are fused through the reconstruction module, and the fused features are upsampled to obtain the coordinated image.
4. The infrared small target detection method based on diffusion model data enhancement as claimed in claim 3 is characterized in that: Before inputting the preprocessed spliced image into the preset image enhancement network, the method further includes: optimizing the preset image enhancement network by using the preset optimized image and the preset enhancement loss function; wherein the optimization process is specifically as follows: Preprocessing the preset optimized image, and inputting the preprocessed preset optimized image into the preset image enhancement network to obtain an optimized coordinated image; According to the optimized coordinated image and the preset optimized image, the preset image enhancement network is optimized by the preset enhancement loss function; wherein the preset enhancement loss function is specifically: in, For said optimized coordinated image, Optimizes the image for the preset in question.
5. The infrared small target detection method based on diffusion model data enhancement as claimed in claim 1, characterized in that: The step of inputting the coordinated image into a preset diffusion model to obtain a resampled image specifically includes: The preset diffusion model includes an encoder, a noise adding module, a denoising network and a decoder; After the coordinated image is input into the preset diffusion model, the coordinated image is encoded into a latent space code by an encoder; Adding noise to the latent space code by the noise adding module to obtain diffuse noise; Resampling the diffuse noise through the denoising network to obtain denoising code; The denoising code is decoded by the decoder to obtain the resampled image.
6. The infrared small target detection method based on diffusion model data enhancement as claimed in claim 5, characterized in that: Before the diffuse noise is resampled by the denoising network, the method further includes: training the preset denoising network by using a preset training image and a preset denoising loss function; wherein the training process is specifically as follows: Encoding the preset training image into a training code by the encoder; Adding noise to the training code by the noise adding module to obtain training noise; According to the training noise, the preset denoising network is trained by the preset denoising loss function; wherein the preset denoising loss function is specifically: in, is Gaussian noise, Encode the training, is the training noise, For the preset denoising network, is the preset time step, Enter the preset conditions. represents a random variable with uniform distribution U(0,1) and t Take the expected value.
7. The infrared small target detection method based on diffusion model data enhancement as claimed in claim 1, characterized in that: The step of performing infrared small target detection on the resampled image through a preset detection network to obtain a detection image specifically includes: The preset detection network includes several residual attention blocks and feature fusion modules; After the resampled image is input into the preset detection network, features of the resampled image are extracted through the plurality of residual attention blocks to obtain a feature pyramid; The feature fusion module performs feature fusion on the feature pyramid to obtain the detection image.
8. The infrared small target detection method based on diffusion model data enhancement as claimed in claim 7, characterized in that: After obtaining the detection image, the method further includes: optimizing the preset detection network according to the detection image and a preset detection loss function; wherein the preset detection loss function is specifically: in, is the detection image, To presume the real result, is the preset smoothing coefficient.
9. An infrared small target detection device based on diffusion model data enhancement, characterized in that: include: Splicing module, enhancement module, diffusion module and detection module; The stitching module is used to convert the image to be detected into a stitched image through Mosaic data enhancement, and preprocess the stitched image; The enhancement module is used to input the preprocessed spliced image into a preset image enhancement network to obtain a coordinated image; The diffusion module is used to input the coordinated image into a preset diffusion model to obtain a resampled image; The detection module is used to perform infrared small target detection on the resampled image through a preset detection network to obtain a detection image.
10. The infrared small target detection device based on diffusion model data enhancement according to claim 9, characterized in that: The enhancement module is used to input the preprocessed stitched image into a preset image enhancement network to obtain a coordinated image, which specifically includes: The preset image enhancement network includes a shallow convolution module, a residual module and a reconstruction module; After the preprocessed spliced image is input into the preset image enhancement network, shallow features are extracted by the shallow convolution module, and deep features are extracted by the residual module; The shallow features and the deep features are fused through the reconstruction module, and the fused features are upsampled to obtain the coordinated image.