A method, apparatus, device, and storage medium for intelligent infrared small target detection

Abstract

This invention relates to the field of infrared small target detection technology, specifically to an intelligent infrared small target detection method, device, equipment, and storage medium. Addressing the problems of boundary irregularities in existing detection technologies and performance degradation caused by blurred edges of small targets in complex scenes, this invention provides an intelligent infrared small target detection method, device, equipment, and storage medium. This method incorporates edge information of small targets using ECMC-Net and suppresses the transmission of redundant information by acquiring semantic information of the small targets. Utilizing the complementarity of edge features and semantic features, it can refine the edges of small targets in cluttered and complex scenes, effectively solving the problem of clutter near the target and performance degradation. Its detection performance is superior to detection methods such as ALCNet and DNANet.

Description

Technical Field

[0001] This invention relates to the field of infrared small target detection technology, specifically to an intelligent infrared small target detection method, device, equipment, and storage medium. Background Technology

[0002] Infrared small target detection technology, with its advantages of strong penetration and anti-interference capabilities, is widely used in target early warning, maritime surveillance, and traffic management. In real-world detection environments, target backgrounds are often very complex, and due to the long imaging distance, targets occupy few pixels, have low grayscale values, and lack color and texture information. Infrared small target detection technology can achieve target detection based on the characteristics of both the target and the background.

[0003] Currently, small object detection algorithms are mainly divided into traditional detection algorithms and deep learning-based detection algorithms. Traditional detection algorithms mainly rely on various filtering methods based on prior assumptions about the background or target, such as LCM, MPCM, and LIG. These methods assume the target is salient and segment it, but this assumption is often too idealistic. In real-world scenarios, prominent edges and target-like features can interfere with detection. Subsequently, methods to separate the background and target were proposed, such as the IPI model and PSTNN model, which transformed the detection problem into low-rank sparse decomposition subunits. These traditional methods perform well in simple scenes, but they are sensitive to noise and perform poorly in scenes with complex backgrounds and dim targets.

[0004] Subsequently, many neural network-based infrared small target detection methods were proposed. MDvsFAcGan proposed a deep adversarial learning framework that uses two opposing sub-units to balance false positives and false negatives. ALCNet designed a feature map cyclic shifting scheme to modularize the local contrast measurement method and also designed an attention modulation module to highlight and preserve small target features. However, ALCNet ignores global feature extraction when using local contrast information, which limits its detection performance in complex background scenes. DNANet designed a densely nested interaction module to realize the interaction between high-level and low-level features, thereby maintaining the feature information of small targets. AGPCNet proposed focusing on the correlation between pixels in patches and achieving target feature representation by fusing information from multiple scales.

[0005] However, due to the small scale of infrared small targets, the network will lose some feature information of the small targets, resulting in feature loss. This leads to the problem of unclear boundaries of small targets in existing detection technologies. In complex scenes, infrared small target detection suffers from the problem of blurred edges of small targets, which leads to a decrease in detection performance. Summary of the Invention

[0006] The purpose of this invention is to address the problems of boundary confusion for small targets in existing detection technologies, and the performance degradation caused by blurred edges of small targets in infrared small target detection in complex scenes. This invention provides an intelligent infrared small target detection method, device, equipment, and storage medium. The method uses ECMC-Net to introduce the edge information of the target, and leverages the complementarity of edge features and semantic features to refine the edges of small targets in cluttered and complex scenes. This intelligent infrared small target detection method obtains the semantic information of small targets through a feature extraction and filtering unit, suppressing the transmission of redundant information and effectively solving the problem of clutter confusion near the target and the resulting performance degradation.

[0007] The technical solution of this application is:

[0008] 1. An intelligent infrared small target detection method: The infrared image data to be tested is input into an edge-combined multi-scale feature network (ECMC-Net) for infrared small target detection. ECMC-Net includes an edge extraction module, a target segmentation module, and a feature fusion module. The ECMC-Net infrared small target detection process includes the following steps:

[0009] A1: The edge extraction module employs five convolutional layers to extract small target edge information from the infrared image data under test at full resolution; the edge extraction module outputs small target edge features F based on the extracted small target edge information. E ;

[0010] A2: The target segmentation module includes a six-layer backbone network unit and a feature extraction and filtering unit. The infrared image data to be tested is preprocessed by two convolutional layers before being input into the target segmentation module.

[0011] The target segmentation module uses a six-layer backbone network to extract features from the preprocessed infrared image data to obtain the first feature set of the image. The preprocessed infrared image data outputs a feature map of a specific resolution at each layer of the backbone network unit. Representing the backbone network Feature maps output by the layer;

[0012] The target segmentation module uses a feature extraction and filtering unit to process the first feature map set F. T Extraction and filtering are performed to obtain the second feature map set. represent The corresponding output feature map, the second feature map set G T That is, the semantic information of small targets in the infrared image to be tested;

[0013] A3: The feature fusion module includes a first fusion unit and a second fusion unit.

[0014] The semantic information of the small target obtained from A2 is input into the first fusion unit of the feature fusion module. The feature fusion module uses the first fusion unit to process each feature map in the second feature map set. The second feature map set G is fused together. T The deep feature map, rich in semantic information, is upsampled to the same size as the shallow feature map to obtain the first fused feature.

[0015] A1 obtains the edge features of the small target F. E The second fusion unit of the input feature fusion module is used by the feature fusion module to process the obtained first fused features. With small target edge features F E The fusion is performed to obtain the second fusion feature, out. fuse ;

[0016] ECMC-Net based on the obtained second fusion feature out fuse Output the predicted image.

[0017] 2. An intelligent infrared small target detection device for performing an intelligent infrared small target detection method; the intelligent infrared small target detection device includes:

[0018] Edge extraction module: Used to extract edge information of small targets and output the edge features F of the small targets. E The infrared image data is input into the edge extraction module of the device. This module employs five convolutional layers to extract small target edge information from the infrared image at full resolution, thus preserving more details. Based on the extracted small target edge information, the edge extraction module outputs the small target edge feature F. E ;

[0019] Target segmentation module: Used to extract semantic information of small targets and output a semantic feature set G of small targets. T The target segmentation module includes a backbone network unit and a feature extraction and filtering unit. Infrared image data outputs a feature map at a corresponding resolution from each of the six layers of the backbone network. After the backbone network outputs feature maps at different resolutions, these feature maps are passed through the feature extraction and filtering unit to extract semantic information about small infrared targets, outputting a semantic feature set G of small targets composed of feature maps at different scales. T ;

[0020] Feature fusion module: Used to fuse the edge features F from the edge extraction module. E The small target semantic feature set G output by the target segmentation module TThe system outputs a predicted image of the small infrared target. The feature fusion module includes a first fusion unit and a second fusion unit. The first fusion unit is used to process the semantic feature set G of the small target. T To perform fusion, the deep feature map is upsampled to the same size as the shallow feature map to obtain the first fused feature. The first fusion feature is then fused using the second fusion unit. With small target edge features F E The fusion is performed to obtain the second fusion feature, out. fuse ECMC-Net based on the second fusion feature out fuse Output the predicted image.

[0021] 3. An electronic device, comprising:

[0022] One or more processors;

[0023] Storage device for storing one or more programs.

[0024] When the one or more programs are executed by the one or more processors, the one or more processors execute the above-described intelligent infrared small target detection method.

[0025] 4. A computer storage medium, characterized in that the storage medium stores at least one instruction, the at least one instruction being loaded and executed by a processor to implement the intelligent infrared small target detection method as described in any one of claims 1 to 8.

[0026] Compared with the prior art, this application has the following advantages:

[0027] The feature extraction section of the intelligent infrared small target detection method and device of this invention consists of two modules. The edge detection module focuses on extracting edge information at full resolution, which helps preserve all target information. The target segmentation module focuses on extracting semantic information from the semantic information network. The feature extraction filtering unit in the target segmentation module can adaptively control the propagation of beneficial information to the next branch, allowing features from the previous branch to be adaptively passed to the next branch. This identifies relevant contextual features to a certain extent while suppressing redundant features. This module helps suppress clutter near small targets and accurately segment infrared small targets. The feature fusion module utilizes the complementarity of edge features and semantic features to refine the edges of small targets in cluttered scenes, fusing edge features and semantic features to suppress the transmission of redundant information. This effectively solves the problem of clutter near targets and reduced detection performance. Attached Figure Description

[0028] Figure 1 This is a three-dimensional schematic diagram of the infrared small target image of the present invention;

[0029] Figure 2 This is a schematic diagram of the ECMC-Net process of the present invention;

[0030] Figure 3 This is a schematic diagram of the ECMC-Net structure of the present invention;

[0031] Figure 4 This is a schematic diagram of the feature extraction and filtering unit structure in the ECMC-Net of this invention;

[0032] Figure 5 This is a comparison chart of the detection performance of the ECMC-Net detection method of this invention with other methods;

[0033] Figure 6 This is a comparison chart of the detection performance of the ECMC-Net detection method of this invention with other methods. Detailed Implementation

[0034] Specific implementation method one: Combining Figures 1 to 6 This embodiment explains that, since small infrared targets lack information such as color and texture, but edges are a scale-invariant feature, the grayscale values ​​of infrared images change abruptly at the boundary between the target and the background. A network can be used to extract edge features, providing the necessary fine target edges for deeper, high-level feature maps. Therefore, developing edge-feature-based target detection methods to refine small target boundaries and improve the detection capability in complex scenes is of great significance for addressing the problem of small infrared target detection in complex environments.

[0035] A method for detecting small infrared targets includes the following steps:

[0036] The infrared image data to be tested is input into ECMC-Net for infrared small target detection. ECMC-Net includes an edge extraction module, a target segmentation module, and a feature fusion module. The process of infrared small target detection by ECMC-Net includes the following steps:

[0037] A1: The edge extraction module employs five convolutional layers to extract small target edge information from the infrared image data under test at full resolution; the edge extraction module outputs small target edge features F based on the extracted small target edge information. E ;

[0038] A2: The target segmentation module consists of a six-layer backbone network unit and a feature extraction and filtering unit;

[0039] A six-layer backbone network is used to extract features from the infrared image data under test. Each layer of the backbone network outputs a feature map of a certain resolution, thus obtaining the first feature map set of the infrared image under test. Representing the backbone network Feature maps output by the layer;

[0040] A feature extraction and filtering unit is used to process the first feature map set F. T Extraction and filtering are performed to obtain the second feature map set. represent The corresponding output feature map, the second feature map set G T That is, the semantic information of small targets in the infrared image to be tested;

[0041] A3: The feature fusion module includes a first fusion unit and a second fusion unit.

[0042] The semantic information of the small target obtained from A2 is input into the first fusion unit of the feature fusion module. The feature fusion module uses the first fusion unit to process each feature map in the second feature map set. The second feature map set G is fused together. T The deep feature map, rich in semantic information, is upsampled to the same size as the shallow feature map to obtain the first fused feature.

[0043] A1 obtains the edge features of the small target F. E The second fusion unit of the input feature fusion module is used by the feature fusion module to process the obtained first fused features. With small target edge features F E The fusion is performed to obtain the second fusion feature, out. fuse ;

[0044] ECMC-Net based on the obtained second fusion feature out fuse Output the predicted image.

[0045] Those skilled in the art can input infrared image data into a network using the steps of this invention, and use an edge extraction module to focus on extracting edge information of small targets: taking advantage of the small size of small targets, the edges of small targets are extracted at full resolution to retain more details and improve the accuracy of small target detection; a target segmentation module is used to extract target semantic information: feature maps of different sizes are generated in different layers of the network, and then the semantic information of small targets at different resolutions is obtained through a feature extraction filtering unit; a feature fusion module is used to couple the semantic information and edge features at different resolutions together for target detection, and output a predicted image. This invention provides an intelligent infrared small target detection method (ECMC-Net). This method, by introducing the edge information of the target and utilizing the complementarity of edge features and semantic features, suppresses the transmission of redundant information, effectively solving the problem of clutter near the target and decreased detection performance. In cluttered and complex scenes, it can refine the edges of small targets, and its detection performance is superior to detection methods such as ALCNet and DNANet.

[0046] Specific Implementation Method Two: Combining Figures 1 to 6 This embodiment describes an intelligent infrared small target detection method: The edge extraction module described in A1 uses a binary cross-entropy loss function as the first loss function, and simultaneously uses Canny edge detection to extract ground truth edge information from the labels. The ECMC-Net edge extraction module is then trained using this method. The first loss function is... BCE The formula is as follows:

[0047]

[0048] In the formula: y i This represents the label of sample i, with 1 for positive and 0 for negative. i This represents the probability that sample i is predicted to be of the positive class, enabling the trained edge extraction module to focus on extracting small target edge information and outputting small target edge features F. E This allows for better preservation of more authentic edge information.

[0049] ECMC-Net uses loss BCE The first loss function described in the second loss function is used to train the image prediction step through the second fused feature output. BCE The following formula is used for processing:

[0050]

[0051] Where y represents the label of the sample and p represents the prediction result, the trained ECMC-Net outputs a more accurate predicted image based on the obtained second fusion feature. Other steps are the same as in Specific Implementation Method 1.

[0052] Specific implementation method three: Combining Figures 1 to 6 This embodiment describes an intelligent infrared small target detection method: the feature extraction and filtering unit described in A2 applies the first feature map set... Extraction and filtering includes the following steps: First feature map set Each first sub-feature map Input the feature extraction and filtering units respectively:

[0053] S1: The input first sub-feature map The first feature map is processed through three branches, each containing a 1×1 convolutional layer and a dilated convolutional subunit. The dilation rates of the dilated convolutional subunits in the three branches are different. In each branch, the first feature map is first processed through a 1×1 convolutional layer to obtain the preliminary branch result.

[0054] S2: The preliminary branch result corresponding to the first branch is processed by the dilated convolution subunit of the first branch to obtain the adjusted branch result of the first branch.

[0055] S3: The adjusted branch result of the first branch is processed by the threshold extraction subunit, and the processed result is added to the preliminary branch result corresponding to the second branch. The added result is processed by the dilated convolution subunit of the second branch to obtain the adjusted branch result of the second branch.

[0056] S4: The adjusted branch result of the second branch is processed by the threshold extraction subunit, and the processed result is added to the preliminary branch result corresponding to the third branch. The added result is processed by the dilated convolution subunit of the third branch to obtain the adjusted branch result of the third branch.

[0057] S5: Concatenate the adjusted branch results of the three branches and output them as the second sub-feature map. All second sub-feature maps The second feature set G is obtained by aggregation. T .

[0058] Simple multi-layer dilated convolution is insufficient to highlight the features required by the object. The semantic features of infrared targets may be submerged due to similar backgrounds or small size. In order to control the transmission of redundant information, a threshold-based feature extraction mechanism is adopted, which effectively controls the transmission of redundant information, filters out irrelevant features, and emphasizes prominent objects in the scene. This unit helps to suppress clutter near small targets, accurately segment infrared small targets, and effectively solves the problem of clutter chaos near targets and decreased detection performance. Other steps are the same as any one of the specific implementation methods one to two.

[0059] Specific implementation method four: Combination Figures 1 to 6This embodiment describes an intelligent infrared small target detection method: the dilated convolutional subunit is processed in the following manner: the dilation rates of the convolutional layers of the dilated convolutional subunits corresponding to the first to third branches are 1, 3, and 5, respectively. The input result of the corresponding branch is first processed through a 3×3 dilated convolutional layer of the corresponding branch, and the processed result is then added to the input result to obtain the adjusted branch result. Other steps are the same as any one of the specific embodiments one to three.

[0060] Specific Implementation Method Five: Combining Figures 1 to 6 This embodiment describes an intelligent infrared small target detection method: the threshold extraction subunit processes data in the following manner:

[0061] B1: The input results are first processed through a 3×3 convolutional layer;

[0062] B2: The processed result is then processed using the Sigmoid function;

[0063] B3: The result of the operation is then multiplied point-by-point with the input result to obtain the output of the threshold extraction subunit. The threshold extraction subunit adds an information transmission channel between adjacent dilated convolution branches. The features of the previous branch are first filtered through a gating function, and useful features are transmitted to the next dilated convolution branch, so that the features of the previous branch can be adaptively passed to the next branch. Through the gating mechanism, the adaptive mechanism can identify relevant contextual features to a certain extent, while suppressing redundant features, thus propagating important information, filtering out irrelevant features, and emphasizing prominent objects in the scene. This effectively solves the problem of clutter near the target and the resulting performance degradation. Other steps are the same as any one of the specific implementation methods one to four.

[0064] Specific Implementation Method Six: Combination Figures 1 to 6 This embodiment describes an intelligent infrared small target detection method: In A3, the first fusion unit adopts a top-down information fusion method, and processes the data according to the following formula:

[0065]

[0066] In the formula: conv(*) represents pointwise convolution operation, Trans(*) represents convolution operation, the purpose of which is to change the number of channels of the feature, and up(*) is bilinear interpolation operation. The point-by-point addition represents the corresponding characteristics. G represents T Shallow and intermediate layer feature maps G represents TThe deep feature map is upsampled to the same size as the shallow feature map, so that the deep feature map and the shallow feature map are gradually merged. Other steps are the same as any one of the specific implementation methods one to five.

[0067] Specific implementation method seven: Combination Figures 1 to 6 This embodiment describes an intelligent infrared small target detection method: the second fusion unit described in A3 integrates the first fusion features. With small target edge features F E Perform fusion into a second fusion feature out fuse Process according to the following formula:

[0068]

[0069] In the formula: Sigmoid represents the Sigmoid function. The point-by-point addition represents the corresponding characteristics. Pointwise multiplication representing corresponding features Emphasizing the salient characteristics of small goals, F E Emphasis is placed on the contour information of small targets. The second fusion unit integrates the first fusion features. By fusing with edge features, the complementarity of shallow features, deep features and edge features is fully utilized, which improves feature accuracy. Other steps are the same as any one of the specific implementation methods one to six.

[0070] Specific Implementation Method Eight: The present invention also provides an intelligent infrared small target detection device: In this embodiment, an intelligent infrared small target detection device includes:

[0071] Edge extraction module: Used to extract edge information of small targets and output the edge features F of the small targets. E The infrared image data is input into the edge extraction module of the device. This module employs five convolutional layers to extract small target edge information from the infrared image at full resolution, thus preserving more details. Based on the extracted small target edge information, the edge extraction module outputs the small target edge feature F. E ;

[0072] Target segmentation module: Used to extract semantic information of small targets and output a semantic feature set G of small targets. T The target segmentation module includes a backbone network unit and a feature extraction and filtering unit. Preprocessed infrared image data outputs a feature map of corresponding resolution at each layer of the six-layer backbone network of the target segmentation module. After the backbone network outputs feature maps of different resolutions, the feature maps are passed through the feature extraction and filtering unit to extract the semantic information of small infrared targets, outputting a small target semantic feature set G composed of feature maps of different scales. T ;

[0073] Feature fusion module: Used to fuse the edge features F from the edge extraction module. E The small target semantic feature set G output by the target segmentation module T The system outputs a predicted image of the small infrared target. The feature fusion module includes a first fusion unit and a second fusion unit. The first fusion unit is used to process the semantic feature set G of the small target. T To perform fusion, the deep feature map is upsampled to the same size as the shallow feature map to obtain the first fused feature. The first fusion feature is then fused using the second fusion unit. With small target edge features F E The fusion is performed to obtain the second fusion feature, out. fuse ECMC-Net obtains the second fusion feature based on out fuse Output the predicted image.

[0074] Detailed Implementation Method Nine: The present invention also provides an electronic device, including:

[0075] One or more processors;

[0076] Storage device for storing one or more programs.

[0077] It should be understood that the device, including any device comprising a processor and a memory as described in this invention, may also include other units or modules that perform display, interaction, processing, control, and other functions via signals or instructions; the memory stores at least one instruction, which is loaded and executed by the processor to implement the intelligent infrared small target detection method described in any one of the specific embodiments 1-8.

[0078] Specific Embodiment Ten: The present invention also provides a computer-readable storage medium, which may be included in the device / apparatus / device described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / device. The computer-readable storage medium carries one or more programs, which, when executed, implement an intelligent infrared small target detection method according to an embodiment of the present invention. According to an embodiment of the present invention, the computer-readable storage medium may be a non-volatile computer-readable storage medium, such as including but not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In the present invention, the computer-readable storage medium may be any tangible medium containing or storing a program, which may be used by or in conjunction with an instruction execution device, apparatus, or apparatus.

[0079] The above description is merely of preferred embodiments of the present invention. It should be understood that the present invention is not limited to the specific embodiments described above. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent substitutions, and improvements made to the above embodiments without departing from the scope of the present invention, based on the technical essence of the present invention, and within the spirit and principles of the present invention, shall still fall within the protection scope of the present invention.

Claims

1. A smart infrared small target detection method, characterized in that, The infrared image data to be detected is input into an edge combination multi-scale feature network for infrared small target detection, the edge combination multi-scale feature network is abbreviated as ECMC-Net, the ECMC-Net comprises an edge extraction module, a target segmentation module and a feature fusion module, and the infrared small target detection process of the ECMC-Net comprises the following steps: A1: The edge extraction module extracts features of the to-be-tested infrared image data by using five convolutional layers; small target edge information of the to-be-tested infrared image data is extracted under full resolution, and the edge extraction module outputs small target edge features according to the extracted small target edge information ; A2: The target segmentation module comprises a six-layer backbone network unit and a feature extraction filtering unit, and the infrared image data to be detected is preprocessed by two convolutional layers before being input into the target segmentation module; The target segmentation module adopts a six-layer backbone network unit to extract features of the preprocessed to-be-tested infrared image data, to obtain a first feature map set of the to-be-tested image The preprocessed to-be-tested infrared image data outputs a feature map with a resolution at each layer of the backbone network of the backbone network unit The feature map represents a feature map output by the i-th layer of the backbone network ​ The target segmentation module adopts a feature extraction filtering unit to extract and filter the first feature map set to obtain a second feature map set , represent corresponding output feature maps, and the second feature map set is the small target semantic information of the to-be-tested infrared image; A3: The feature fusion module comprises a first fusion unit and a second fusion unit, The small target semantic information obtained by A2 is input into a feature fusion module of the ECMC-Net, the feature fusion module uses a first fusion unit to fuse each feature map in a second feature map set to obtain a first fusion feature The deep feature map with rich semantic information in the second feature map set is up-sampled to the same size as the shallow feature map to obtain a first fusion feature ; Obtaining small target edge features from A1 The feature fusion module of the ECMC-Net, the feature fusion module uses a second fusion unit to fuse the obtained first fusion features With small target edge features Fusion to obtain second fusion features ; The ECMC-Net obtains second fusion features according to the obtained first fusion features An output prediction image is output.

2. The intelligent infrared small target detection method according to claim 1, characterized in that: The ECMC-Net is pre-trained, and in the training process, the edge extraction module in A1 adopts a binary cross-entropy loss function As the first loss function, the true value of the edge information extracted from the label by using the canny edge detection is used to train the edge extraction module of the ECMC-Net, and the first loss function The formula is as follows: wherein: denotes the label of the sample with 1 for positive class and 0 for negative class, denotes the label of the sample predicted to be positive class; The ECMC-Net adopts As the second loss function, the second fusion feature The output prediction image step is trained, and the second loss function The following formula is used for processing: Wherein, y represents the label of the sample, and p represents the prediction result.

3. The intelligent infrared small target detection method according to claim 1 or 2, characterized in that: The feature extraction filtering unit described in A2 filters the first feature map set The feature extraction filtering in B includes the following steps: the first feature map set Each sub-feature map in B The feature extraction filtering unit is respectively input S1: inputting the input sub-feature map respectively pass through three branches, the expansion rates of the corresponding hollow convolution sub-units of the three branches are different from each other, and in each branch, the first feature map is first processed by a 1x1 convolution layer to obtain a preliminary branch result; S2: The preliminary branch result corresponding to the first branch is processed through the first branch of the hollow convolution subunit to obtain the adjusted branch result of the first branch; S3: The adjusted branch result of the first branch is processed through the threshold extraction subunit, and the processed result is added to the preliminary branch result corresponding to the second branch, and the added result is processed through the hollow convolution subunit of the second branch to obtain the adjusted branch result of the second branch; S4: The adjusted branch result of the second branch is processed through the threshold extraction subunit, and the processed result is added to the preliminary branch result corresponding to the third branch, and the added result is processed through the hollow convolution subunit of the third branch to obtain the adjusted branch result of the third branch; S5: splice the three branch results after adjustment, and output as a second feature map .

4. The intelligent infrared small target detection method according to claim 3, characterized in that: The hollow convolution subunit processes in the following manner: the convolution layer expansion rates of the hollow convolution subunits corresponding to the first to third branches are 1, 3 and 5 respectively, the input result of the corresponding branch is first processed through a 3x3 hollow convolution layer of the corresponding branch, and then the processed result is added to the input result to obtain the adjusted branch result of the corresponding branch.

5. The intelligent infrared small target detection method according to claim 4, characterized in that: The threshold extraction subunit processes in the following manner: B1: the input result is first processed through a 3x3 convolution layer; B2: the processed result is operated through a Sigmoid function; B3: the operated result is point-by-point multiplied with the input result to obtain the output result of the threshold extraction subunit.

6. The intelligent infrared small target detection method according to claim 1, characterized in that: The first fusion unit in A3 adopts a top-down information fusion manner and processes according to the following formula: In the formula: represents a point-wise convolution operation, represents a convolution operation, which changes the channel number of the feature, is a bilinear interpolation operation, represents a point-wise addition of the corresponding features, represents a shallow feature map, represents a deep feature map.

7. The intelligent infrared small target detection method according to claim 1 or 6, characterized in that: The second fusion unit described in A3 fuses the first fusion feature with the small target edge feature into a second fusion feature according to the following formula: wherein: represents a Sigmoid function, represents point-wise addition of the respective features, represents point-wise multiplication of the respective features.

8. An intelligent infrared small target detection device, characterized in that: An intelligent infrared small target detection device comprises: The edge extraction module is configured to extract small target edge information of the small target, and output small target edge features ; input the infrared image data into an edge extraction module of the device; the edge extraction module adopts five convolution layers to extract small target edge information of the to-be-tested infrared image at full resolution, so as to retain more details; and the edge extraction module outputs small target edge features according to the extracted small target edge information ; The target segmentation module is configured to extract semantic information of the small target and output a small target semantic feature set The target segmentation module comprises a backbone network unit and a feature extraction filtering unit. The preprocessed infrared image data outputs a corresponding resolution feature map at each layer of the six-layer backbone network of the target segmentation module. After the backbone network outputs feature maps of different resolutions, the feature extraction filtering unit is used to extract semantic information of the small infrared target, and a small target semantic feature set composed of feature maps of different scales is output ; The feature fusion module is configured to fuse the edge features of the edge extraction module and the small target semantic feature set output by the target segmentation module , and output a predicted image of the infrared small target. The feature fusion module comprises a first fusion unit and a second fusion unit. The first fusion unit is configured to fuse the small target semantic feature set , up-sample the deep feature map to the same size as the shallow feature map, and obtain first fusion features . The second fusion unit is configured to fuse the first fusion features with the small target edge features , and obtain second fusion features . The edge combined multi-scale feature network ECMC-Net outputs the predicted image according to the second fusion features .

9. An electronic device, comprising: comprises: one or more processors; a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors execute the intelligent infrared small target detection method according to any one of claims 1 to 8.

10. A computer storage medium, characterized in that, The storage medium stores at least one instruction, and the at least one instruction is loaded and executed by the processor to realize the intelligent infrared small target detection method according to any one of claims 1 to 8.

Images