Infrared small target detection model establishment method and detection method based on multi-scale attention feature superposition

Abstract

The present invention discloses a method for establishing an infrared small target detection model and a detection method based on multi-scale attention feature superposition, belonging to the field of infrared small target detection. The model establishment method includes: obtaining a set of infrared images containing infrared small targets, annotating the target frame containing the small targets, as well as the target pixels and background pixels in each image; using the annotated infrared image set to train an initial detection network to obtain an infrared small target detection model; in the initial detection network, a multi-scale region candidate network is used to extract features of different scales in the input image, extract the spatial attention of each scale feature, and then fuse the features of different scales together based on the spatial attention to thereby screen out a target frame containing the infrared small target; after a semantic extractor obtains the semantic features of the target frame, a segmentation encoder identifies the probability of each pixel belonging to the small target, thereby obtaining an infrared small target detection result. The present invention can improve the detection quality of infrared small targets.

Description

Technical Field

[0001] The present invention belongs to the field of infrared small target detection, and more specifically, relates to an infrared small target detection model establishment method and a detection method based on multi-scale attention feature superposition. Background Art

[0002] Infrared imaging plays an important role in search and rescue, detection, early warning, and aerospace guidance due to its unrestricted imaging range and long imaging distance. However, due to the long imaging distance of infrared images, such as several kilometers, the targets in the image are generally only a few pixels or even a single pixel. Therefore, the targets in infrared images are generally called small infrared targets or weak targets. Detecting weak targets in infrared images has always been a widely used task. In practical applications, the task of detecting small infrared targets faces the following difficulties:

[0003] (1) The size of infrared small targets is very small, often only a few pixels or even only one pixel, which makes many detection methods unsuitable for infrared small target detection.

[0004] (2) The brightness of small infrared targets is relatively low. If the image background (the background in the sky, such as the sun or other heat sources) is brighter, the target may be submerged in it, making the target detection task difficult.

[0005] (3) The contrast between the target and the background area of ​​the infrared image is low, and there is a lot of noise in the entire image. The weak target to be detected has low brightness and small size, so it is easy to be confused with the noise and difficult to detect.

[0006] In order to cope with the task of infrared small target detection, there are two main processing methods. The first is the traditional method based on the shallow features of the target. This method uses the structural features of small targets in infrared images as a detection method. Methods such as maximum median filtering, feature kernel transformation, top-hat morphological filtering, and local contrast enhancement can obtain the spatial characteristics of small targets in infrared images and achieve small target detection. However, traditional detection methods based on shallow target features have certain shortcomings. Because traditional methods can only rely on the low-dimensional features of the target and cannot obtain the deep features of the target, they are prone to feature confusion. In today's situation where the resolution of infrared images is getting higher and higher and the number of countermeasures is increasing, it is difficult to adapt to all scenarios, and the detection effect is often poor.

[0007] Another approach to infrared small target detection is based on deep learning. Deep learning algorithms are self-learning and feature extraction at multiple levels, making them widely used in the field of infrared small targets. In particular, the introduction of algorithms such as ACMNet, DNANet, ALCNet, and ResNet has paved the way for deep learning in infrared small target detection. Deep learning-based detection methods typically first extract image features using a typical deep learning network such as ResNet. Then, an image segmentation encoder is used to detect infrared small targets based on these features. However, using deep learning methods to detect infrared small targets still presents certain challenges. Deep learning methods typically process images using convolutional and pooling layers to extract high-dimensional features. However, small and weak targets in infrared images have low contrast and small size, making their features extremely indistinct. These features can be lost in background or interference. Therefore, directly using typical deep learning networks to extract features can result in poor detection quality.

[0008] In general, the existing infrared small target detection methods cannot adapt to the characteristics of small infrared targets such as small size, low brightness, and low contrast with the background, and the detection quality needs to be further improved. Summary of the Invention

[0009] In response to the defects of the existing technology and the need for improvement, the present invention provides a method for establishing an infrared small target detection model and a detection method based on the superposition of multi-scale attention features. Its purpose is to improve the detection model structure based on the characteristics of small infrared targets, such as small size, low brightness, and low contrast with the background, so as to improve the detection quality of small infrared targets.

[0010] To achieve the above objectives, according to one aspect of the present invention, a method for establishing an infrared small target detection model based on multi-scale attention feature superposition is provided, comprising:

[0011] Obtain a set of infrared images containing small infrared targets, mark the target box where each small target is located in each infrared image, and mark the pixels belonging to the small target and the background separately;

[0012] An initial detection network is established to detect small infrared targets in infrared images. The labeled infrared image set is used to train the initial detection network. After training, an infrared small target detection model is obtained.

[0013] Among them, the initial detection network includes a multi-scale region proposal network, a semantic extractor, and a segmentation encoder;

[0014] The multi-scale region candidate network includes a multi-scale attention feature extraction network and a feature mapping convolution layer. The multi-scale attention feature extraction network is used to extract features of different scales in the input image. After extracting the spatial attention of the features at each scale, the features of different scales are fused together based on the spatial attention to obtain a fused feature map. The feature mapping convolution layer is used to map the fused feature map into a five-channel feature map and filter out the target box containing the small infrared target. In the five-channel feature map, the information at each point includes the center coordinates, length, width, and confidence level of the corresponding target box.

[0015] The semantic extractor is used to extract the semantic features of the input image and extract the semantic features belonging to the screened target box to obtain the ROI image block;

[0016] The segmentation encoder is used to encode the ROI image block to identify the probability of each pixel belonging to a small target and obtain the infrared small target detection result.

[0017] Furthermore, the multi-scale attention feature extraction network includes: a plurality of enhanced attention extraction modules connected in sequence, each enhanced attention extraction module is further connected to a downsampling layer;

[0018] The enhanced attention extraction module includes: a splitting layer, four branches, a connection layer, and a residual structure; the splitting layer is used to split the feature map into four sub-feature maps with equal number of channels, which are input into four branches respectively; the four branches are used to extract features from the four sub-feature maps respectively; the connection layer is used to stack the features extracted from the four branches by channel to obtain multi-scale attention features; the residual structure is used to fuse the multi-scale attention features with the image to obtain a fused feature map;

[0019] Among them, the first branch includes: feature extraction unit C1, 1×1 convolution layer and addition unit A1 in sequence; the second branch includes: feature extraction unit C2, 3×3 convolution layer, multiplication unit M1, feature enhancement unit E1 and addition unit A2 in sequence; the third branch includes: feature extraction unit C3, 5×5 convolution layer, feature extraction unit C4, multiplication unit M3, feature enhancement unit E2 and addition unit A3 in sequence; the fourth branch includes: feature extraction unit C5, 7×7 convolution layer, feature extraction unit C6, feature extraction unit C7, multiplication unit M3, feature enhancement unit E3; the spatial attention module SA1 is also included between the first branch and the second branch, the spatial attention module SA2 is also included between the second branch and the third branch, and the spatial attention module SA3 is also included between the third branch and the fourth branch;

[0020] The 1×1 convolution layer takes the fusion result of the output features of feature extraction units C1 and C2 as input, and obtains feature F1 through convolution operation. The 3×3 convolution layer takes the fusion result of the output features of feature extraction units C2 and C3 as input, and obtains feature F2 through convolution operation. The 5×5 convolution layer takes the fusion result of the output features of feature extraction units C3 and C4 as input, and obtains feature F3 through convolution operation.

[0021] The spatial attention module SA1 takes the output of feature F1 as input; the multiplication unit M1 takes the spatial attention map extracted by the spatial attention module SA1 and the feature F2 as input, and multiplies the two to obtain feature FA1; the feature enhancement unit E1 takes feature FA1 as input, and enhances it to obtain feature FAE1; the feature extraction unit C4 takes the result of the fusion of features F2 and F3 as input, and obtains feature F after convolution operation. 23 ; Feature extraction unit C6 takes the result of fusion of feature F2 and feature F4 output by 7×7 convolution as input, and obtains feature F after convolution operation. 34 ;

[0022] The addition unit A1 takes the features F1 and FAE1 as input, and adds them together to get the features extracted by the first branch; the spatial attention module SA2 takes the features FAE1 as input; the multiplication unit M2 takes the spatial attention map extracted by the spatial attention module SA2 and the feature F 23 The feature enhancement unit E2 takes the feature FA2 as input and obtains the feature FAE2 after convolution operation; the feature extraction unit C7 takes the feature F as input and obtains the feature FAE2 after convolution operation; 23 and F 34 The fused result is the input, and after the convolution operation, the feature F is obtained. 234 ;

[0023] The addition unit A2 takes the features FAE1 and FAE2 as input, and adds them together to obtain the features extracted by the second branch; the spatial attention module SA3 takes the feature FAE2 as input; the multiplication unit M3 takes the spatial attention map extracted by the spatial attention module SA3 and the feature F 234 As input, the feature FA3 is obtained by multiplying the two together; the feature enhancement unit E3 takes the feature FA3 as input, enhances it to obtain the feature FAE3, and the feature FAE3 is also used as the feature extracted by the fourth branch;

[0024] The addition unit A3 takes features FAE2 and FAE3 as input and adds them together to obtain the features extracted by the third branch.

[0025] Furthermore, when training the initial detection network using the annotated infrared image set, a two-stage training approach is adopted;

[0026] The first phase of training includes: training the initial detection network. During the training process, the parameters of the segmentation encoder are frozen, and the parameters of the multi-scale region candidate network are updated using the target box annotation results as supervision information;

[0027] The second stage of training includes: after the first stage of training, unlocking the segmentation encoder, training the initial detection network, and using the target box annotation results and pixel annotation results as supervision information to update the parameters of the multi-scale region candidate network and the segmentation encoder.

[0028] Furthermore, in the first stage training, the training loss includes the target box prediction loss of the multi-scale region proposal network; in the second stage training, the training loss includes the target box prediction loss of the multi-scale region proposal network and the segmentation loss of the segmentation encoder;

[0029] Among them, the target box prediction loss is used to measure the loss of the filtered target box for the target box annotation result; the segmentation loss is used to measure the loss of the segmentation result relative to the pixel value annotation result.

[0030] Furthermore, the expression of the target box prediction loss is:

[0031]

[0032] Among them, L RPN Represents the target box prediction loss; M represents the total number of filtered target boxes; Indicates the confidence of the target frame; c represents the true value of the confidence. If the predicted target frame is the target frame with the marked infrared small target, its true value of the confidence is 1, otherwise it is 0; L CIoU Represents the CIoU between the predicted target box and the reference anchor box.

[0033] Furthermore, the expression of segmentation loss is:

[0034]

[0035] Among them, L c represents the segmentation loss; N represents the total number of pixels; Indicates the probability prediction result that the i-th pixel belongs to the infrared small target, p i Represents the true value of the labeled category.

[0036] Furthermore, the target frame containing the infrared small target is filtered out, including:

[0037] Eliminate target boxes whose confidence is lower than the preset confidence threshold.

[0038] Furthermore, filtering out the target frame containing the small infrared target further includes at least one of the following operations:

[0039] Eliminate target boxes whose area is larger than the preset size threshold;

[0040] Eliminate target frames whose aspect ratios are not within the preset aspect ratio range.

[0041] According to another aspect of the present invention, there is provided a method for detecting small infrared targets, comprising:

[0042] Input the infrared image to be detected into the pre-established infrared small target detection model;

[0043] Among them, the infrared small target detection model is established by the above-mentioned infrared small target detection model establishment method based on multi-scale attention feature superposition provided by the present invention.

[0044] According to another aspect of the present invention, a computer-readable storage medium is provided, comprising a stored computer program; when the computer program is executed by a processor, the device where the computer-readable storage medium is located is controlled to execute the above-mentioned method for establishing an infrared small target detection model based on multi-scale attention feature superposition provided by the present invention, and / or the above-mentioned infrared small target detection method provided by the present invention.

[0045] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects:

[0046] (1) The infrared small target detection model based on multi-scale attention feature superposition established by the present invention, in which the multi-scale region candidate network specifically extracts features of different scales of the input image when extracting infrared image features, and extracts the spatial attention of features of each scale, and fuses features of different scales together based on spatial attention. Through this feature extraction method, the attention map can be used to enhance the small target features, so that the weak small target features are retained in the image, overcoming the disadvantage of information loss of infrared small targets in feature transmission. This feature extraction method is adapted to the characteristics of small infrared targets, such as small size, low brightness, and low contrast with the background. The extracted features are of high quality, can achieve effective infrared small target detection, and improve the detection quality of infrared small targets.

[0047] (2) In the preferred embodiment of the present invention, the multi-scale attention feature extraction network in the multi-scale region candidate network specifically includes four branches, which respectively extract features from the sub-feature maps obtained by splitting the feature map, and then splice them as the overall features; the feature scales extracted by the four branches are different, which can effectively adapt to the variable infrared target sizes, and the attention of the smaller-scale features will be gradually transferred to the larger-scale features, realizing the feature enhancement of weak targets, and allowing the features of weak targets to be retained in the image, effectively improving the quality of the extracted features, and thus improving the quality of the infrared small target detection results.

[0048] (3) In the preferred embodiment of the present invention, a two-stage training method is specifically adopted when training the model. In the first stage of training, only the parameters of the multi-scale region candidate network are updated. In the second stage of training, the parameters of the multi-scale region candidate network and the segmentation encoder are updated at the same time. Since the segmentation result of the segmentation encoder depends on the prediction result of the multi-scale region candidate network for the target frame, the present invention is based on this two-stage training method. When the multi-scale region candidate network has a certain prediction ability through the first stage of training, the parameters of the segmentation encoder are updated, thereby ensuring the effectiveness of the parameter update and improving the training efficiency.

[0049] (4) The size of small infrared targets is usually small and they all have relatively regular shapes. In the preferred embodiment of the present invention, on the basis of screening the target frame based on the confidence level, the target frame with a large area or abnormal aspect ratio will be eliminated, thereby improving the accuracy of the target frame prediction and further improving the accuracy of the subsequent segmentation prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 A schematic diagram of a method for establishing an infrared small target detection model based on multi-scale attention feature superposition and a detection method provided by an embodiment of the present invention;

[0051] Figure 2 A schematic diagram of the structure of an infrared small target detection model based on multi-scale attention feature superposition provided by an embodiment of the present invention;

[0052] Figure 3 Schematic diagram of a multi-scale attention feature extraction network provided by an embodiment of the present invention;

[0053] Figure 4 A schematic diagram of infrared small target detection results provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0054] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0055] In the present invention, the terms "first", "second", etc. (if any) in the present invention and the drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0056] In order to solve the technical problem of poor detection quality of existing infrared small target detection methods, the present invention provides a method for establishing an infrared small target detection model and a detection method based on the superposition of multi-scale attention features. The overall idea is to improve the feature extraction part in the detection model based on the characteristics of small infrared targets, such as small size, low brightness, and low contrast with the background, so that it can process infrared weak targets of different sizes, and use the attention map to enhance the small target features so that the weak target features can be retained in the image, thereby achieving effective small target detection, overcoming the disadvantage of information loss of weak targets in feature transmission, and ultimately improving the detection quality of small infrared targets.

[0057] The following are examples.

[0058] Example 1:

[0059] A method for establishing an infrared small target detection model based on multi-scale attention feature superposition, such as Figure 1 As shown, this embodiment includes three main steps: a dataset construction step, a network construction step, and a training step. The dataset construction step is used to annotate infrared images containing small infrared targets to construct a dataset that can be used to train the infrared small target model. The network construction step is used to build a network model that can accurately detect infrared small targets. The training step uses the dataset constructed in the dataset construction step to train the network model constructed in the network construction step, so that the established network model has the ability to perform actual infrared small target detection tasks. Each goal is explained in detail below.

[0060] In this embodiment, the data set construction step specifically includes: obtaining an infrared image set containing infrared small targets, marking the target frame where each small target is located in each infrared image, and marking the pixels belonging to the small target and the background respectively.

[0061] It's easy to understand that the typical size of small infrared targets ranges from 1 to 10 pixels. In this embodiment, the size of small infrared targets is also 1 to 10 pixels, and the main type of small infrared targets is drones. If the target is too large, general methods can be used for detection rather than specific algorithms.

[0062] In the network model established in this embodiment, images undergo four downsampling cycles before being forwarded through the network. Therefore, the infrared images acquired in this embodiment are cropped to a pixel count where both the length and width are multiples of 16. Furthermore, considering that most infrared images of weak targets are distant, small, and have few pixels, the number of pixels in the infrared images acquired in this embodiment is limited to between 128 and 384 for image application scenarios.

[0063] In addition, in order to give the deep learning algorithm sufficient feature fitting, this embodiment uses at least 200 typical infrared small target images in different scenes for labeling, training, and recognition to ensure the robustness of the algorithm in different scenes.

[0064] For each image in the infrared image collection, this embodiment adopts a two-stage supervised annotation method, including reference position annotation and target segmentation annotation. The reference position annotation is to mark the target box where each small target is located, and the target segmentation annotation is to mark the pixels belonging to the small target and the background separately.

[0065] In this embodiment, the reference position annotation covers the position of each small infrared target. The minimum bounding rectangle of the small infrared target is annotated using the coordinates based on the target box. In actual applications, the data is recorded in the CSV format commonly used by various operating systems. The box format is:

[0066] I m =[B1,B2,...,B n ]

[0067] Where, I m represents the mth infrared image, B i Represents the Box coordinate frame of the i-th infrared small target contained in the image, where the position of each coordinate frame is defined as follows:

[0068] Box=[x l ,y l ,x r ,y r ]

[0069] Where x l ,y l Respectively represent the upper left corner coordinates of the infrared small target, x r ,y r Indicates the coordinates of the lower right corner of the target.

[0070] Target segmentation annotation is pixel-level annotation, which is used to monitor the accuracy of the final small target segmentation detection image at the second-level output of the network. To effectively distinguish each location and reduce the amount of data, the annotation method uses image format. During image annotation, an 8-bit single-channel grayscale image of the same size as the infrared image is generated to record the exact pixel location of the target. In this annotated image, the pixel value 255 is used to represent the true value location of the target, and the pixel value 0 is used to represent the background area location. That is, in this image:

[0071]

[0072] Where I represents the entire annotated image, p irepresents the pixel value of the i-th pixel in the image, o i Indicates whether the pixel position belongs to the target or the background. If its value is 1, it means it belongs to the target. If the value is 0, it means it is the background.

[0073] Based on the above-mentioned annotation method, the pixel value can simultaneously represent the pixel category. Specifically, a pixel value of 255 indicates that the pixel is a target pixel, and a pixel value of 0 indicates that the pixel is a background pixel.

[0074] The above segmentation and annotation data is used as the second-stage output supervision data of the network, which is also the final output data of the entire network, that is, the detection results of infrared small targets are represented by binary segmentation maps.

[0075] In order to adapt to the characteristics of small infrared targets, such as small size, long distance, low brightness, and low contrast with the background, in the network building step of this embodiment, the network model built, that is, the initial detection network, has the following structure: Figure 2 As shown, it includes a multi-scale region proposal network, a semantic extractor, and a segmentation encoder;

[0076] The multi-scale region candidate network includes a multi-scale attention feature extraction network and a feature mapping convolution layer. The multi-scale attention feature extraction network is used to extract features of different scales in the input image. After extracting the spatial attention of the features at each scale, the features of different scales are fused together based on the spatial attention to obtain a fused feature map. The feature mapping convolution layer is used to map the fused feature map into a five-channel feature map and filter out the target box containing the small infrared target. In the five-channel feature map, the information at each point includes the center coordinates, length, width, and confidence level of the corresponding target box.

[0077] The semantic extractor is used to extract the semantic features of the input image and extract the semantic features belonging to the screened target box to obtain the ROI image block;

[0078] The segmentation encoder is used to encode the ROI image block to identify the probability of each pixel belonging to a small target and obtain the infrared small target detection result.

[0079] In the above-mentioned model constructed in this embodiment, when extracting infrared image features, the multi-scale region candidate network specifically extracts features of different scales of the input image, and extracts the spatial attention of the features of each scale, and fuses the features of different scales together based on the spatial attention. Through this feature extraction method, the attention map can be used to enhance the features of small targets, so that the weak target features are retained in the image, overcoming the disadvantage of information loss of small infrared targets in feature transmission. This feature extraction method is adapted to the characteristics of small infrared targets, such as small size, low brightness, and low contrast with the background. The extracted features are of high quality, can achieve effective detection of small infrared targets, and improve the detection quality of small infrared targets.

[0080] Specifically, in this embodiment, the multi-scale attention feature extraction network includes: four enhanced attention extraction modules connected in sequence, namely Figure 2 In the EAModule, each enhanced attention extraction module is also connected to a downsampling layer. Optionally, in this embodiment, the downsampling layer is specifically connected to the pooling layer (i.e. Figure 2 Pool) implementation after EAModule in.

[0081] The structure of the enhanced attention extraction module is as follows Figure 3 As shown, it specifically includes: split layer, four branches, connection layer and residual structure;

[0082] The splitting layer is used to split the feature map into 4 sub-feature maps with equal number of channels, which are input into four branches respectively; I represents the channel input of the input image X of the enhanced attention extraction module, then the channel input of each sub-feature map is C I / 4; For the first EAModule, its input image X is the input image of the entire multi-scale attention feature extraction network; For the remaining EAModules, its input image X is the feature map after downsampling of the features output by the previous level EAModule;

[0083] The four branches are used to extract features from the four sub-feature maps respectively;

[0084] The connection layer is used to stack the features extracted from the four branches by channel to obtain multi-scale attention features;

[0085] The residual structure is used to fuse multi-scale attention features and images to obtain a fused feature map.

[0086] like Figure 3 As shown in the figure, among the four branches of EAModule, the first branch includes: feature extraction unit C1, 1×1 convolution layer and addition unit A1 in sequence; the second branch includes: feature extraction unit C2, 3×3 convolution layer, multiplication unit M1, feature enhancement unit E1 and addition unit A2 in sequence; the third branch includes: feature extraction unit C3, 5×5 convolution layer, feature extraction unit C4, multiplication unit M3, feature enhancement unit E2 and addition unit A3 in sequence; the fourth branch includes: feature extraction unit C5, 7×7 convolution layer, feature extraction unit C6, feature extraction unit C7, multiplication unit M3, feature enhancement unit E3; the spatial attention module SA1 is also included between the first branch and the second branch, the spatial attention module SA2 is also included between the second branch and the third branch, and the spatial attention module SA3 is also included between the third branch and the fourth branch;

[0087] The 1×1 convolution layer takes the fusion result of the output features of the feature extraction units C1 and C2 as input, and obtains feature F1 through convolution operation. The 3×3 convolution layer takes the fusion result of the output features of the feature extraction units C2 and C3 as input, and obtains feature F2 through convolution operation. The 5×5 convolution layer takes the fusion result of the output features of the feature extraction units C3 and C4 as input, and obtains feature F3 through convolution operation. In this embodiment, after preliminary image feature extraction is performed on each sub-feature map, feature extraction is performed again using convolution kernels with sizes of 1, 3, 5, and 7. By performing feature extraction through these four convolutions, it is ensured that infrared small targets of all scales are covered. At the same time, before feature extraction, the features in the next channel will be fused into the features of the current channel, thereby effectively expanding the receptive field while extracting features of different scales.

[0088] The spatial attention module SA1 takes the output of feature F1 as input; the multiplication unit M1 takes the spatial attention map extracted by the spatial attention module SA1 and the feature F2 as input, and multiplies the two to obtain feature FA1; the feature enhancement unit E1 takes feature FA1 as input, and enhances it to obtain feature FAE1; the feature extraction unit C4 takes the result of the fusion of features F2 and F3 as input, and obtains feature F after convolution operation. 23 ; Feature extraction unit C6 takes the result of fusion of feature F2 and feature F4 output by 7×7 convolution as input, and obtains feature F after convolution operation. 34 ;

[0089] The addition unit A1 takes the features F1 and FAE1 as input, and adds them together to get the features extracted by the first branch; the spatial attention module SA2 takes the features FAE1 as input; the multiplication unit M2 takes the spatial attention map extracted by the spatial attention module SA2 and the feature F 23 The feature enhancement unit E2 takes the feature FA2 as input and obtains the feature FAE2 after convolution operation; the feature extraction unit C7 takes the feature F as input and obtains the feature FAE2 after convolution operation; 23 and F 34 The fused result is the input, and after the convolution operation, the feature F is obtained. 234 ;

[0090] The addition unit A2 takes the features FAE1 and FAE2 as input, and adds them together to obtain the features extracted by the second branch; the spatial attention module SA3 takes the feature FAE2 as input; the multiplication unit M3 takes the spatial attention map extracted by the spatial attention module SA3 and the feature F 234 As input, the feature FA3 is obtained by multiplying the two together; the feature enhancement unit E3 takes the feature FA3 as input, enhances it to obtain the feature FAE3, and the feature FAE3 is also used as the feature extracted by the fourth branch;

[0091] The addition unit A3 takes features FAE2 and FAE3 as input and adds them together to obtain the features extracted by the third branch.

[0092] Based on the above structure, each branch transfers feature information while extracting features of different scales. The attention maps of smaller-scale features are passed down step by step. That is, the attention map of the features extracted by the first branch is gradually transferred to the second, third, and fourth branches, the attention map of the features extracted by the second branch is gradually transferred to the third and fourth branches, and the attention map of the features extracted by the third branch is transferred to the fourth branch. Through this feature information transfer method, the features of weak targets are enhanced and the features of weak targets can be retained in the image, effectively improving the quality of the extracted features. After extracting features of different scales, before extracting features again, the features of the previous branch are fused into the features of the current branch, thereby making the extracted features smoother. Each time the attention map is multiplied by the feature, the feature is enhanced, and the enhanced features are passed back to the previous branch. Through this feature transfer method, the target features can be effectively enhanced and the contrast between the target features and the background features can be effectively increased.

[0093] Optionally, in this embodiment, the feature extraction units are all 3×3 convolutional layers, and the feature enhancement modules are all 3×3 convolutional layers; when the residual structure fuses the multi-scale attention features and the image, it will specifically use 3×3 convolutional layers and 1×1 convolutional layers to perform convolution operations on the multi-scale attention features and the original input image respectively to achieve smoothing processing, and then add the two together to finally obtain the fused feature map Y.

[0094] X represents the tensor image input to EAModule, which can be expressed as follows:

[0095]

[0096] Where H and W represent the height and width of the tensor image X respectively.

[0097] When the tensor image X is input, the EAModule processes the tensor image X as follows:

[0098] First, split it into four sub-feature maps, process different features at different levels, and the number of feature channels of the four sub-feature maps is C I / 4. After extracting the basic image semantics using the feature extraction convolution layer, the number of channels of each sub-feature map is C O / 4, and then use convolution kernels with sizes of 1, 3, 5, and 7 for feature extraction to adapt to different target sizes.

[0099] Use channel fusion connection to connect the different features of each two layers in parallel to achieve multi-scale feature fusion. The results are as follows:

[0100] X imso =[X i ,X i+1 ]

[0101] In the above formula, X imso Represents the multi-scale hierarchical fusion result, X i Represents the features extracted by the convolution operation at the i-th scale. The small-scale spatial attention results of the previous layer are then used to enhance the large-scale features of the next layer, thereby highlighting the target information. The calculation method is as follows:

[0102]

[0103] Where, X i With X i+1 represents the features before processing, X′ i+1 represents the processed features, softmax is the activation function, and SA represents the spatial attention operation.

[0104] After the attention fusion calculation, a layer of enhanced convolution is superimposed to achieve a thorough fusion of features at different scales. After the fusion of one layer of attention features is completed, the addition operation is used to make it complementary to the features of the previous layer. The calculation method is as follows:

[0105]

[0106] Where X′ i Represents the features after superposition and complementation, X i represents the feature layer before processing, Indicates features that have been reinforced by upper-level attention.

[0107] After the feature processing of the four levels is completed, the feature maps of the four discrete sub-channels are fused again using Concat connection. The operation is as follows:

[0108] X'=Concat(X1,X2,X3,X4)∈R C×H×W

[0109] In the formula, X' represents the fused features, X i ∈R C / 4×H×W To fuse the features of each level before, the number of feature channels after fusion becomes C again O .

[0110] In order to achieve feature extraction with a wider dimension, in this embodiment, the output dimensions of the 4-layer enhanced attention extraction modules in the multi-scale attention feature extraction network are set to 64, 64, 128, and 256 respectively. As the network depth increases, the number of channels is gradually increased to obtain more high-dimensional semantics. It should be noted that the setting of the number of enhanced attention extraction modules and the output dimension setting of each enhanced attention extraction module here is only a preferred embodiment and should not be understood as the only limitation of the present invention. In some other embodiments of the present invention, it can also be set to other numbers according to the actual detection task characteristics. Accordingly, the number of downsampling modules can be adjusted accordingly, and the output dimension of each enhanced attention extraction module can also be flexibly adjusted as needed.

[0111] Based on the above structure, the multi-scale attention feature extraction network can extract high-quality 256-channel feature maps. In the five-channel feature maps obtained by the feature mapping convolution layer, each point corresponds to a target box prediction result, which can be expressed as follows:

[0112]

[0113] Among them, They represent the x, y coordinates of the center position of the predicted target box, the length and width of the target box, and the predicted confidence value; the confidence reflects the possibility that the target box prediction is accurate.

[0114] Based on the above five-channel feature map, this embodiment will eliminate the target frame with low confidence and only retain the target frame with high confidence. Specifically, the confidence threshold is set to 0.5, and the target frame with confidence lower than 0.5 is excluded.

[0115] Furthermore, considering that small infrared targets are relatively small in size and that drone targets are all relatively regular in shape, appearing as point targets or point-like area targets in infrared images, in order to accurately filter out target frames containing small infrared targets, this embodiment further sets a size threshold and aspect ratio range. Optionally, the size threshold is 15, and the aspect ratio range is [0.5, 2]. Based on this, target frames with length and width greater than 15 pixels will be further eliminated, and target frames with aspect ratios less than 0.5 or greater than 2 will be eliminated to ensure the rationality and accuracy of the selected target frames. Since all target frames are floating-point numbers after network extraction, when calculating the length and width of the target frame, they are rounded according to the principle of coordinate value +0.5.

[0116] Optionally, in this embodiment, the segmentation encoder is constructed using a Transformer encoder structure, and the input it receives is an image block, which is obtained by obtaining a mask on the semantic image from the ROI extracted by the multi-scale region candidate network. In order to supplement the semantic dimension of the image block, a semantic generator is used to obtain high-dimensional semantics. It is a fully convolutional structure with 7 convolutional layers and no downsampling units. The image size is guaranteed to remain unchanged during the transformation process, and the output feature dimension is 512. When the mask corresponding to the ROI is applied to the high-dimensional semantic feature map, the image block of the suspected target area is obtained, so the dimension of all image blocks is 512, the same size as the ROI.

[0117] After obtaining the target image block, it is serialized and sent to the codec structure. The segmentation encoder structure consists of 4 layers. For an input sequence containing n image blocks, its representation is as follows:

[0118] S={w1,w2,...,w n},w i ∈R d×h×w

[0119] Where w i represents the i-th image patch, and d represents the dimension of the input image patch, which is the same as the dimension of the semantic extractor, i.e., 512.

[0120] The input and output structures of the 4-layer structure are consistent. The input sequence S is projected onto the Q, K, and V matrices, and the matrix dimensions are all d. s ×n,d s Represents the maximum length-width product of the input image block, and its value is 128. For each layer of the encoder, the attention calculation method used is as follows:

[0121]

[0122] in Is a size coefficient used to prevent the weight of the dot product attention from returning to zero. In order to enable the attention structure to regress different semantic features, this embodiment specifically uses a 4-head attention structure and uses a feedforward network to map the attention output, which is represented as follows:

[0123] M(Q,K,V)=Concat(h1,…,h m )W O

[0124] h i =A(Q i ,K i ,V i )

[0125] X'=FFN(M)+M

[0126] In the formula, M(Q,K,V) represents the fusion output of multiple heads, h i represents the output of the i-th head, W O is the weighted matrix, dimension is md s ×d, while FFN represents the feedforward network and X′ represents the sequence of image patches after the attention map, which represents the target probability at each mask position.

[0127] In the training step of this embodiment, when using the infrared image set that has undergone two-stage annotation to train the initial detection network, a two-stage training method is specifically adopted, and the infrared image set is randomly divided into a training set and a validation set in a ratio of 4:1. The image batch size for model training is 8, and the total number of training rounds is 30 rounds;

[0128] The first phase of the two-stage training involves training the initial detection network. During training, the parameters of the segmentation encoder are frozen, and the parameters of the multi-scale region candidate network are updated using the object box annotation results as supervision information. The first phase of training is conducted on the entire training set for 5 rounds. After the first phase of training, the multi-scale region candidate network has a certain ability to predict object boxes.

[0129] The second stage of training includes: after the completion of the first stage of training, unlocking the segmentation encoder, training the initial detection network, and updating the parameters of the multi-scale region candidate network and the segmentation encoder with the target box annotation results and pixel annotation results as supervision information; the learning rate hyperparameter used in the training process gradually decreases from 0.005 to 0.001 as the loss continues to decrease. The multi-scale RPN and image block segmentation encoder are trained together for 25 rounds until the losses of the multi-scale RPN and image block segmentation encoder all converge to within 0.5% of the maximum loss.

[0130] Since the segmentation result of the segmentation encoder depends on the prediction result of the multi-scale region candidate network for the target box, this embodiment is based on this two-stage training method. The parameters of the segmentation encoder are updated only when the multi-scale region candidate network has a certain prediction ability through the first stage training. This ensures the effectiveness of the parameter update and can improve training efficiency.

[0131] In this embodiment, in the first stage of training, the training loss includes the target box prediction loss of the multi-scale region proposal network; in the second stage of training, the training loss includes the target box prediction loss of the multi-scale region proposal network and the segmentation loss of the segmentation encoder;

[0132] Among them, the target box prediction loss is used to measure the loss of the filtered target box for the target box annotation result; the segmentation loss is used to measure the loss of the segmentation result relative to the pixel value annotation result; the expression of the target box prediction loss is:

[0133]

[0134] Among them, L RPN Represents the target box prediction loss; M represents the total number of filtered target boxes; Indicates the confidence of the target frame; c represents the true value of the confidence. If the predicted target frame is the target frame with the marked infrared small target, its true value of the confidence is 1, otherwise it is 0; L CIoU Represents the CIoU (Complete Intersection on Union) between the predicted target box and the annotated target box; because the size of the feature map is smaller than the original image, the coordinate information of each point needs to be mapped proportionally to the original image information. The downsampling level used by the multi-scale region proposal network is 16 times. Considering that the size of the target generally does not exceed 10 pixels, the base anchor box size is set to 10 pixels in this embodiment. The CIoU of the coordinate box prediction value output by each anchor box is calculated as follows:

[0135] L CIoU =1-IoU+R DIoU +αυ

[0136] Where R DIoU Represents DIoU loss, α represents the loss weight parameter, υ represents the similarity of the aspect ratio between the predicted target box and the labeled target box, and IoU represents the similarity between the predicted target box and the labeled target box.

[0137] To facilitate the calculation of the segmentation loss of the segmentation encoder, this embodiment connects a classification head after the segmentation encoder when training the model, and maps the image block predicted by the segmentation encoder to a vector, which is represented as follows:

[0138]

[0139] Where N represents the total number of pixels in the image, Represents the probability prediction value of a single pixel belonging to a small infrared target. Similarly, the segmentation annotation image in the two-stage annotation is also stretched into a corresponding vector to calculate the cross entropy classification loss as the segmentation loss. The expression of the segmentation loss is:

[0140]

[0141] Among them, L c represents the segmentation loss; N represents the total number of pixels; represents the probability prediction result of the i-th pixel belonging to the infrared small target; p i Indicates the true value of the annotated category. According to the annotation results of the pixel value, if the pixel value of the pixel is annotated as 255, it means that the pixel belongs to the infrared small target, and the corresponding category true value is 1. If the pixel value of the pixel is annotated as 0, it means that the pixel belongs to the background, and the corresponding category true value is 0.

[0142] Based on the calculated target box prediction loss and segmentation loss, in the second stage training, the overall loss L all for:

[0143] L all =L c +L RPN

[0144] After the initial detection network undergoes two-stage training, the classification head is removed and it can be used to detect small infrared targets. In general, this embodiment, by designing a multi-scale, multi-level feature extraction network, can effectively adapt to the variable sizes of infrared targets and is suitable for small targets and extremely weak targets. It also uses a multi-level attention structure to achieve feature enhancement of weak targets, thereby ensuring the stability of feature transfer of the detected targets and achieving better detection and positioning effects. The use of two-stage supervision, basic position supervision of small infrared targets and pixel semantic supervision of the final segmentation of the targets, can ensure the effectiveness of feature extraction within the network. Using a pixel-level semantic segmentation map as the infrared small target detection result instead of a simple coordinate box as the output, the position of each pixel of the small infrared target can be accurately located, greatly improving the detection accuracy.

[0145] Example 2:

[0146] An infrared small target detection method, such as Figure 1 Shown, including:

[0147] Input the infrared image to be detected into the pre-established infrared small target detection model;

[0148] Among them, the infrared small target detection model is established by the infrared small target detection model establishment method based on multi-scale attention feature superposition in the above-mentioned embodiment 1.

[0149] Based on the infrared small target detection model established in Example 1 of the present invention, after the infrared image to be detected is input into the model, a two-stage reasoning will be performed. In the first stage, the multi-scale region candidate network will generate a feature map with a size of 1 / 16 of the infrared image to be detected, which is expressed as:

[0150] F∈I C×H×W

[0151] Where C represents the number of channels output by the RPN structure, H and W represent the height and width of the output feature map respectively. The number of channels is set to 5, and each point represents the coordinates and confidence information of a target box:

[0152] Box=[x,y,w,h,conf]

[0153] Where x, y, w, h, and conf represent the center coordinates x and y of the target, the length and width of the target box, and the confidence of the target detection, respectively. Since it is a single-category target detection classification, the one-hot vector of the category division is not considered.

[0154] All target boxes with a confidence level less than 0.5 are filtered out using a confidence threshold of 0.5. Furthermore, target boxes with both length and width greater than 15 pixels and with aspect ratios less than 0.5 or greater than 2 are removed to select reasonable target boxes. When calculating the length and width of the target box, the coordinate value is rounded to +0.5.

[0155] According to the target box detection results, the ROI image block at the corresponding position in the semantic image is extracted, and an 8-bit single-channel mask image with the same size as the input image is generated. The image block position is filled with 255, and the rest of the area is filled with 0. It is extracted at the same time as the image block area for attention encoding.

[0156] After the mask image and ROI image block are extracted, they are sent to the second-stage image segmentation encoder for probability prediction. The encoder outputs the probability that each pixel position in the image block containing the target is a small infrared target. The probability value of the position is ANDed with the mask image to obtain the segmentation results of all small targets in the image.

[0157] The target prediction probability value is a floating-point number with a range of [-1, 1]. The target detection result is mapped to an 8-bit integer single channel using a threshold of 0.5. Specifically, if the predicted probability value is greater than or equal to 0.5, the pixel value is set to 255 to identify it as a pixel in the infrared small target; if the predicted probability value is less than 0.5, the pixel value is set to 0 to identify it as a pixel in the background. Once the identification is completed, the segmentation result is represented by an 8-channel image.

[0158] Figure 4 The figure shows the detection results of small targets detected in different infrared images using the infrared small target detection method provided by this embodiment. For easy viewing, the infrared small targets in the predicted results and the annotated true values ​​are marked with dotted boxes. Figure 4 It can be seen from the results that the infrared small target detection method provided in this embodiment can detect small and weak targets in infrared images relatively accurately.

[0159] The following uses existing detection methods as a comparison, and combines the detection results of different detection methods on public datasets to further analyze and verify the beneficial effects that can be achieved by the present invention. Specifically, four public infrared small target detection datasets, IDSADT, SIRST, NUDT-SIRST, and CGAN, are used. All four datasets contain a rich variety of infrared scenes, such as cities, skies, and mountains. Each dataset contains at least 300 infrared images, all of which reflect the small size of targets in infrared scenes, including low signal-to-noise ratio, low contrast, and a very small number of target pixels. The target characteristics of each dataset are shown in Table 1.

[0160] Table 1 Dataset target feature description

[0161]

[0162] In Table 1, SNR represents signal-to-noise ratio, and Cr represents contrast.

[0163] The existing detection methods selected include LCM, MPCM, ALCNet, ACMNet, and IAANet.

[0164] IoU and nIoU (Normalized IoU) are used as the evaluation indicators of small target detection balance, which represent the intersection of the detection results on a single image and the entire dataset respectively. The calculation method is as follows:

[0165]

[0166] Where N represents the number of images in the test dataset, TP i Indicates the number of pixels in the i-th image where both the detection value and the true value are true, T i , P i They represent the number of true value pixels and the number of pixels of predicted positive examples respectively.

[0167] The IoU and nIoU of the detection results of different methods on each dataset are shown in Table 2.

[0168] Table 2 Detection indicators of different detection methods on various data sets

[0169]

[0170] According to the data shown in Table 2, the infrared small target detection method provided by this embodiment outperforms other existing methods in terms of IoU and nIoU indicators on various data sets.

[0171] Example 3:

[0172] A computer-readable storage medium includes a stored computer program; when the computer program is executed by a processor, the device where the computer-readable storage medium is located is controlled to execute the infrared small target detection model establishment method based on multi-scale attention feature superposition provided in the above-mentioned embodiment 1, and / or the infrared small target detection method provided in the above-mentioned embodiment 2.

[0173] It will be easily understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for establishing an infrared small target detection model based on multi-scale attention feature superposition, characterized in that: include: Obtain a set of infrared images containing small infrared targets, mark the target box where each small target is located in each infrared image, and mark the pixels belonging to the small target and the background separately; Establishing an initial detection network for detecting small infrared targets in infrared images, and training the initial detection network using the labeled infrared image set. After the training is completed, an infrared small target detection model is obtained; The initial detection network includes a multi-scale region proposal network, a semantic extractor and a segmentation encoder; The multi-scale region candidate network includes a multi-scale attention feature extraction network and a feature mapping convolution layer; the multi-scale attention feature extraction network is used to extract features of different scales of the input image, extract the spatial attention of the features of each scale, and then fuse the features of different scales together based on the spatial attention to obtain a fused feature map; the feature mapping convolution layer is used to map the fused feature map into a five-channel feature map, and filter out the target frame containing the small infrared target; in the five-channel feature map, the information at each point includes the center coordinates, length, width and confidence of the corresponding target frame; The semantic extractor is used to extract semantic features of the input image, and extract semantic features belonging to the screened target frame therefrom to obtain an ROI image block; The segmentation encoder is used to encode the ROI image block to identify the probability that each pixel therein belongs to a small target, and obtain an infrared small target detection result.

2. The method for establishing an infrared small target detection model based on multi-scale attention feature superposition according to claim 1, characterized in that: The multi-scale attention feature extraction network includes: a plurality of enhanced attention extraction modules connected in sequence, each enhanced attention extraction module is further connected to a downsampling layer; The enhanced attention extraction module includes: a splitting layer, four branches, a connection layer, and a residual structure; the splitting layer is used to split the feature map into four sub-feature maps with equal number of channels, which are input into four branches respectively; the four branches are used to extract features from the four sub-feature maps respectively; the connection layer is used to stack the features extracted from the four branches by channel to obtain multi-scale attention features; the residual structure is used to fuse the multi-scale attention features with the image to obtain a fused feature map; Among them, the four branches, the first branch includes: feature extraction unit C1, 1×1 convolution layer and addition unit A1 in sequence; the second branch includes: feature extraction unit C2, 3×3 convolution layer, multiplication unit M1, feature enhancement unit E1 and addition unit A2 in sequence; the third branch includes: feature extraction unit C3, 5×5 convolution layer, feature extraction unit C4, multiplication unit M3, feature enhancement unit E2 and addition unit A3 in sequence; the fourth branch includes: feature extraction unit C5, 7×7 convolution layer, feature extraction unit C6, feature extraction unit C7, multiplication unit M3, feature enhancement unit E3; the first branch and the second branch also include a spatial attention module SA1, the second branch and the third branch also include a spatial attention module SA2, and the third branch and the fourth branch also include a spatial attention module SA3; The 1×1 convolution layer takes the fusion result of the output features of feature extraction units C1 and C2 as input, and obtains feature F1 through convolution operation. The 3×3 convolution layer takes the fusion result of the output features of feature extraction units C2 and C3 as input, and obtains feature F2 through convolution operation. The 5×5 convolution layer takes the fusion result of the output features of feature extraction units C3 and C4 as input, and obtains feature F3 through convolution operation. The spatial attention module SA1 takes the output of feature F1 as input; the multiplication unit M1 takes the spatial attention map extracted by the spatial attention module SA1 and the feature F2 as input, and multiplies the two to obtain feature FA1; the feature enhancement unit E1 takes feature FA1 as input, and enhances it to obtain feature FAE1; the feature extraction unit C4 takes the result of the fusion of features F2 and F3 as input, and obtains feature F after convolution operation. 23 ; Feature extraction unit C6 takes the result of fusion of feature F2 and feature F4 output by 7×7 convolution as input, and obtains feature F after convolution operation. 34 ; The addition unit A1 takes the features F1 and FAE1 as input, and adds them together to get the features extracted by the first branch; the spatial attention module SA2 takes the features FAE1 as input; the multiplication unit M2 takes the spatial attention map extracted by the spatial attention module SA2 and the feature F 23 The feature enhancement unit E2 takes the feature FA2 as input and obtains the feature FAE2 after convolution operation; the feature extraction unit C7 takes the feature F as input and obtains the feature FAE2 after convolution operation; 23 and F 34 The fused result is the input, and after the convolution operation, the feature F is obtained. 234 ; The addition unit A2 takes the features FAE1 and FAE2 as input, and adds them together to obtain the features extracted by the second branch; the spatial attention module SA3 takes the feature FAE2 as input; the multiplication unit M3 takes the spatial attention map extracted by the spatial attention module SA3 and the feature F 234 As input, the feature FA3 is obtained by multiplying the two together; the feature enhancement unit E3 takes the feature FA3 as input, enhances it to obtain the feature FAE3, and the feature FAE3 is also used as the feature extracted by the fourth branch; The addition unit A3 takes features FAE2 and FAE3 as input and adds them together to obtain the features extracted by the third branch.

3. The method for establishing an infrared small target detection model based on multi-scale attention feature superposition according to claim 1 or 2, characterized in that: When the initial detection network is trained using the annotated infrared image set, a two-stage training is adopted; The first phase of training includes: training the initial detection network. During the training process, the parameters of the segmentation encoder are frozen, and the parameters of the multi-scale region candidate network are updated using the target box annotation results as supervision information; The second stage of training includes: after the first stage of training, unlocking the segmentation encoder, training the initial detection network, and using the target box annotation results and pixel annotation results as supervision information to update the parameters of the multi-scale region candidate network and the segmentation encoder.

4. The method for establishing an infrared small target detection model based on multi-scale attention feature superposition according to claim 3, characterized in that: In the first stage of training, the training loss includes the target box prediction loss of the multi-scale region proposal network; in the second stage of training, the training loss includes the target box prediction loss of the multi-scale region proposal network and the segmentation loss of the segmentation encoder; Among them, the target box prediction loss is used to measure the loss of the filtered target box for the target box annotation result; the segmentation loss is used to measure the loss of the segmentation result relative to the pixel value annotation result.

5. The method for establishing an infrared small target detection model based on multi-scale attention feature superposition according to claim 4, characterized in that: The expression of the target box prediction loss is: Among them, L RPN Represents the target box prediction loss; M represents the total number of filtered target boxes; Indicates the confidence of the target frame; c represents the true value of the confidence. If the predicted target frame is the target frame with the marked infrared small target, its true value of the confidence is 1, otherwise it is 0; L CIoU Represents the CIoU between the predicted target box and the annotated target box.

6. The method for establishing an infrared small target detection model based on multi-scale attention feature superposition according to claim 4, characterized in that: The expression of the segmentation loss is: Among them, L c represents the segmentation loss; N represents the total number of pixels; Indicates the probability prediction result that the i-th pixel belongs to the infrared small target, p i Represents the true value of the labeled category.

7. The method for establishing an infrared small target detection model based on multi-scale attention superposition according to claim 1 or 2, characterized in that: The step of filtering out a target frame containing a small infrared target includes: Eliminate target boxes whose confidence is lower than the preset confidence threshold.

8. The method for establishing an infrared small target detection model based on multi-scale attention feature superposition according to claim 7, characterized in that: The step of filtering out the target frame containing the small infrared target further includes at least one of the following operations: Eliminate target boxes whose area is larger than the preset size threshold; Eliminate target frames whose aspect ratios are not within the preset aspect ratio range.

9. A method for detecting small infrared targets, characterized in that: include: Input the infrared image to be detected into the pre-established infrared small target detection model; The infrared small target detection model is established by the infrared small target detection model establishment method based on multi-scale attention feature superposition according to any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that Including a stored computer program; when the computer program is executed by a processor, it controls the device where the computer-readable storage medium is located to execute the method for establishing an infrared small target detection model based on multi-scale attention feature superposition as described in any one of claims 1 to 8, and / or the infrared small target detection method as described in claim 9.

Images