Infrared small target image detection method, system, storage medium and device

Abstract

The embodiment of the present invention discloses an infrared small target image detection method, comprising: constructing an infrared small target detection model, including a multi-layer convolution fusion module, a multi-receptive field fusion module, an efficient channel attention module and a fully connected layer module; extracting and fusing features of the original infrared image through the multi-layer convolution fusion module to obtain a multi-layer convolution fusion feature map; extracting and fusing features of several receptive fields of the multi-layer convolution fusion feature map through the multi-receptive field fusion module to obtain a multi-receptive field fusion feature map; adaptively adjusting the channel weights of the multi-receptive field fusion feature map through the efficient channel attention module to obtain an adaptively adjusted feature map; mapping the adaptively adjusted feature map through the fully connected layer module to obtain the semantic segmentation result of the original infrared image and use it as the detected infrared small target image. The present invention improves the accuracy of image detection through the processing of different modules.

Description

Technical Field

[0001] The present invention relates to the technical field of small target detection and recognition, and in particular to an infrared small target image detection method, system, storage medium and device. Background Art

[0002] Infrared small target detection is an important research direction in the field of computer vision, and plays an irreplaceable role in application scenarios such as military reconnaissance, night monitoring and disaster warning. Due to their small size, low signal-to-noise ratio and susceptibility to complex background interference, traditional detection methods face challenges such as low detection accuracy and high computational complexity.

[0003] Traditional infrared dim target detection methods, such as median filtering, top-hat transform, and maximum mean / maximum median filtering, mainly suppress background noise through spatial domain or transform domain filtering technology, thereby enhancing the target signal; however, these methods are often difficult to effectively distinguish the target from the background in complex and changeable infrared scenes, such as when there is strong thermal radiation interference, non-uniform background and low contrast environment, resulting in poor detection effect and low computational efficiency.

[0004] With the vigorous development of deep learning technology, especially the widespread application of convolutional neural networks (CNN) in the field of image recognition, new ideas have been provided for infrared dim small target detection. Deep learning models such as MDvsFA-cGAN and FCRN have achieved improved target detection performance in complex backgrounds by introducing generative adversarial networks (GAN) and fully convolutional network structures. These methods use the powerful feature learning ability of deep networks to automatically extract high-level abstract features from data, thereby overcoming the limitations of traditional methods to a certain extent.

[0005] However, current deep learning methods still have shortcomings in infrared dim small target detection: on the one hand, since small targets in infrared images are usually small in size, the pooling operation in the deep network and the excessively large receptive field may lead to the loss of small target information and affect the detection accuracy; on the other hand, although the channel attention mechanism can enhance the model's attention to important features, the complex attention module design also increases the computational complexity and training difficulty of the model. Summary of the invention

[0006] Based on this, it is necessary to propose an infrared dim target image detection method to address the above problems.

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

[0008] Constructing an infrared small target detection model, wherein the infrared small target detection network includes a multi-layer convolution fusion module, a multi-receptive field fusion module, an efficient channel attention module and a fully connected layer module;

[0009] The multi-layer convolution fusion module is used to extract and fuse the original infrared image to obtain a multi-layer convolution fusion feature map;

[0010] The multi-receptive field fusion module extracts and fuses the features of several receptive fields of the multi-layer convolution fusion feature map to obtain a multi-receptive field fusion feature map;

[0011] Adaptively adjusting the channel weights of the multi-receptive field fusion feature map through the efficient channel attention module to obtain an adaptively adjusted feature map;

[0012] The adaptively adjusted feature map is mapped through the fully connected layer module to obtain the semantic segmentation result of the original infrared image and use it as the detected infrared weak target image.

[0013] In the above scheme, the feature extraction and fusion of the original infrared image by the multi-layer convolution fusion module to obtain the multi-layer convolution fusion feature map specifically includes:

[0014] In each convolution node of the multi-layer convolution fusion module, any original infrared image Through the convolution layer, normalization layer, activation function and residual connection operation in sequence:

[0015] F conv =ReLU(BN(Conv(I in )));

[0016] Among them, the residual connection operation expression is: F out =F conv +I in ;

[0017] Each convolution node is a repeated structure of convolution layer, normalization layer, activation function and residual connection. When there are N nodes, the multi-layer convolution fusion feature map F multi for:

[0018]

[0019] in, is the output feature map of the Nth node.

[0020] In the above scheme, the multi-receptive field fusion module extracts and fuses the features of several receptive fields of the multi-layer convolution fusion feature map to obtain the multi-receptive field fusion feature map, which specifically includes:

[0021] According to the convolutional layers with residual learning in three different receptive fields, the target characteristics in different local areas are obtained, wherein the convolution kernel sizes corresponding to the three receptive fields are 3, 5, and 7 respectively, and the corresponding depth features are obtained using receptive fields of different sizes:

[0022]

[0023]

[0024] Among them, x is the input, G i_j is the value of the corresponding convolution kernel, i is 1-3, j is 1 or 2, F i is the feature obtained after a series of operations;

[0025] Concatenate the three features according to the first dimension to obtain the multi-receptive field fusion feature map F output : F output =F1+F2+F3

[0026] Among them, F i It is the feature obtained after a series of operations, and the value of i is 1-3.

[0027] In the above scheme, the multi-receptive field fusion feature map F is obtained output Afterwards, the method further comprises:

[0028] The multi-receptive field fusion feature map F output With the original infrared image I in Perform pixel multiplication and compare the multiplication result with the original infrared image I in Addition;

[0029] The added feature map is processed through the convolution layer to obtain the feature map

[0030] F in =Conv(I in +(I in ×F output ))

[0031] Among them, I in is the original infrared image, F output It is a multi-receptive field fusion feature map.

[0032] In the above scheme, the adaptive adjustment of the channel weights of the multi-receptive field fusion feature map by the efficient channel attention module to obtain the adaptively adjusted feature map specifically includes:

[0033] The feature map F inPerform global average pooling to obtain aggregate features:

[0034] F conv =Conv1d(F in );

[0035] The channel weights are generated by a fast 1D convolution of size k, where k is adaptively determined according to the mapping of the channel dimension C:

[0036]

[0037] Where |t| odd represents the odd number closest to t, γ and b are both 1;

[0038] The obtained feature map is normalized by SIGMOD and then fused with the multi-receptive field feature map F output Multiply to generate adaptive adjustment feature map

[0039] F CAM =σ(F conv )×F output

[0040] Among them, F output is the multi-receptive field fusion feature map, F conv is the aggregation feature.

[0041] In the above scheme, after mapping the adaptively adjusted feature map through the fully connected layer module, obtaining the semantic segmentation result of the original infrared image and using it as the detected infrared dim small target image, the method further includes:

[0042] The optimal loss function is obtained based on scale-sensitive loss and positioning-sensitive loss:

[0043]

[0044] in, and They represent scale-sensitive loss and positioning-sensitive loss respectively;

[0045] The scale-sensitive loss expression is:

[0046]

[0047]

[0048] where Var(·,·) is the function that obtains the variance of the provided scalar, |A p |with|A gt The larger the gap between |, the smaller w, the greater the scale sensitivity loss, A pFor a given set of predicted images, A gt is a set of truth values.

[0049] In the above solution, obtaining the positioning sensitivity loss includes:

[0050] The given prediction image A p The set and the truth value are like A gt The coordinates of all pixels in the set are averaged to obtain A p and A gt The corresponding first center point and second center point are denoted as c p =(x p ,y p ) and c gt =(x gt ,y gt );

[0051] Convert the coordinates of the first center point and the second center point into the corresponding distance d in the polar coordinate system p and angle θ p :

[0052]

[0053]

[0054] Among them, d p is the corresponding distance in the polar coordinate system, θ p is the corresponding angle θ in the polar coordinate system p ;

[0055] Determine the positioning sensitive loss based on the predicted center point and the true center point of the target:

[0056]

[0057] Among them, d gt and θ gt c gt Position-sensitive loss and angular position loss in polar coordinates.

[0058] The present application also proposes an infrared small target image detection system, the system comprising: a detection model building unit and a detection result determination unit;

[0059] The detection model construction unit is used to construct an infrared small target detection model, and the infrared small target detection network includes a multi-layer convolution fusion module, a multi-receptive field fusion module, an efficient channel attention module and a fully connected layer module;

[0060] The detection result determination unit is used to perform feature extraction and fusion on the original infrared image through the multi-layer convolution fusion module to obtain a multi-layer convolution fusion feature map; perform feature extraction and fusion of several receptive fields on the multi-layer convolution fusion feature map through the multi-receptive field fusion module to obtain a multi-receptive field fusion feature map; perform adaptive adjustment of channel weights on the multi-receptive field fusion feature map through the efficient channel attention module to obtain an adaptively adjusted feature map; map the adaptively adjusted feature map through the fully connected layer module to obtain a semantic segmentation result of the original infrared image and use it as a detected infrared weak target image.

[0061] The present application also proposes a readable storage medium storing a computer program. When the computer program is executed by a processor, the processor executes the following steps:

[0062] Constructing an infrared small target detection model, wherein the infrared small target detection network includes a multi-layer convolution fusion module, a multi-receptive field fusion module, an efficient channel attention module and a fully connected layer module;

[0063] The multi-layer convolution fusion module is used to extract and fuse the original infrared image to obtain a multi-layer convolution fusion feature map;

[0064] The multi-receptive field fusion module extracts and fuses the features of several receptive fields of the multi-layer convolution fusion feature map to obtain a multi-receptive field fusion feature map;

[0065] Adaptively adjusting the channel weights of the multi-receptive field fusion feature map through the efficient channel attention module to obtain an adaptively adjusted feature map;

[0066] The adaptively adjusted feature map is mapped through the fully connected layer module to obtain the semantic segmentation result of the original infrared image and use it as the detected infrared weak target image.

[0067] The present application also proposes a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor executes the following steps of the computer program:

[0068] Constructing an infrared small target detection model, wherein the infrared small target detection network includes a multi-layer convolution fusion module, a multi-receptive field fusion module, an efficient channel attention module and a fully connected layer module;

[0069] The multi-layer convolution fusion module is used to extract and fuse the original infrared image to obtain a multi-layer convolution fusion feature map;

[0070] The multi-receptive field fusion module extracts and fuses the features of several receptive fields of the multi-layer convolution fusion feature map to obtain a multi-receptive field fusion feature map;

[0071] Adaptively adjusting the channel weights of the multi-receptive field fusion feature map through the efficient channel attention module to obtain an adaptively adjusted feature map;

[0072] The adaptively adjusted feature map is mapped through the fully connected layer module to obtain the semantic segmentation result of the original infrared image and use it as the detected infrared weak target image.

[0073] The embodiment of the present invention has the following beneficial effects: firstly, an infrared small target detection model is constructed, and the infrared small target detection network includes a multi-layer convolution fusion module, a multi-receptive field fusion module, an efficient channel attention module and a fully connected layer module; the original infrared image is feature extracted and fused by the multi-layer convolution fusion module to obtain a multi-layer convolution fusion feature map; the multi-receptive field fusion module extracts and fuses the features of several receptive fields of the multi-layer convolution fusion feature map to obtain a multi-receptive field fusion feature map; the multi-receptive field fusion feature map is adaptively adjusted in terms of channel weights by the efficient channel attention module to obtain an adaptively adjusted feature map; the adaptively adjusted feature map is mapped by the fully connected layer module to obtain a semantic segmentation result of the original infrared image and use it as a detected infrared small target image. In this method, each module is optimized for the key issues of infrared small target detection. The combination of multi-layer convolution fusion, multi-receptive field fusion and efficient channel attention modules enables the network to perform well in extracting features, capturing multi-scale information and optimizing feature weights. Through the collaborative work of the above modules, the scheme can efficiently detect small targets from complex infrared backgrounds, which is conducive to improving detection efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0075] in:

[0076] Figure 1 A schematic diagram of a flow chart of an infrared small target image detection method in an embodiment;

[0077] Figure 2 A schematic diagram of the structure of an infrared small target detection model in an embodiment;

[0078] Figure 3 Based on Figure 2 Schematic diagram of multi-layer convolution fusion module;

[0079] Figure 4 Based on Figure 2 Schematic diagram of the multi-receptive field fusion module. DETAILED DESCRIPTION

[0080] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0081] In the following description, a large number of specific details are given in order to provide a more thorough understanding of the present invention; however, it is obvious to those skilled in the art that the present invention can be implemented without one or more of these details; in other examples, in order to avoid confusion with the present invention, some technical features known in the art are not described, and it should be understood that the present invention can be implemented in different forms and should not be interpreted as limited to the embodiments presented here; on the contrary, providing these embodiments will make the disclosure thorough and complete and fully convey the scope of the present invention to those skilled in the art.

[0082] The purpose of the terms used herein is only to describe specific embodiments and is not intended to be limiting of the present invention. When used herein, the singular forms "a", "an" and "said / the" are also intended to include the plural forms, unless the context clearly indicates otherwise, and it should also be understood that the terms "consisting of" and / or "comprising", when used in this specification, determine the presence of the features, integers, steps, operations, elements and / or parts, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, parts and / or groups. When used herein, the term "and / or" includes any and all combinations of the relevant listed items.

[0083] In order to thoroughly understand the present invention, a detailed structure will be presented in the following description to illustrate the technical solution proposed by the present invention; optional embodiments of the present invention are described in detail as follows, but in addition to these detailed descriptions, the present invention may also have other implementation methods.

[0084] like Figure 1 As shown, in one embodiment, a method for detecting a small infrared target image is provided. The method for detecting a small infrared target image includes steps S101 to S105, which are described in detail as follows:

[0085] S101. Construct an infrared small target detection model. The infrared small target detection network includes a multi-layer convolution fusion module, a multi-receptive field fusion module, an efficient channel attention module and a fully connected layer module.

[0086] like Figure 2 The figure shows a schematic diagram of the structure of the infrared small target detection model. The infrared small target detection model takes the infrared image as input. The lower layer is the main part of the model. The multi-layer convolution module combined with the residual module is used to fuse the target information extracted from each layer. The number of channels of the feature map is converted into the final number of categories through a specific convolution layer. The features extracted from different receptive field sizes are fused through the multi-receptive field fusion module. The fused result is pixel-wise multiplied with the input image of the upper layer, and the result is added to the input image. The purpose of these two operations is to enhance the correlation between the feature and the original input. Then, an efficient channel attention module is used to further strengthen the attention to the key features of small targets in the infrared image. Finally, the output of the network is mapped to the semantic segmentation result of the original input image through the full convolution module, that is, the small target information is predicted, and the detected infrared image binarization map is output.

[0087] S102, extracting and fusing features of the original infrared image through a multi-layer convolution fusion module to obtain a multi-layer convolution fusion feature map;

[0088] The multi-layer convolution structure can gradually extract features at different levels in infrared images, from edges and textures to higher-level semantic information. By fusing feature maps at different levels, it can combine shallow detail information and deep semantic information to improve the model's sensitivity to weak targets.

[0089] like Figure 3 The figure shows a schematic diagram of a multi-layer convolution fusion module. Compared with the traditional Unet architecture, the present invention adds multiple nodes between the encoder and the decoder. The number of downsampling layers of the encoder is controlled at two layers to ensure that small targets will not be lost during the downsampling process. In the figure, f represents a feature image, a black dotted arrow represents a convolution process, a green dotted arrow represents a downsampling process, and a black solid arrow represents a convolution and upsampling process. Indicates addition processing.

[0090] In some embodiments, the original infrared image is subjected to feature extraction and fusion by a multi-layer convolution fusion module to obtain a multi-layer convolution fusion feature map, which specifically includes:

[0091] In each convolution node of the multi-layer convolution fusion module, any original infrared image Through the convolution layer, normalization layer, activation function and residual connection operation in sequence:

[0092] Fconv =ReLU(BN(Conv(I in )));

[0093] Preferably, in the process of feature extraction, in order to maintain the connection between the original input features and the extracted features, a residual connection is introduced. This connection method can help the model learn more context information and alleviate the problem of gradient disappearance: Wherein, the residual connection operation expression is: F out =F conv +I in ;

[0094] Each convolution node is a repeated structure of convolution layer, normalization layer, activation function and residual connection. When there are N nodes, the multi-layer convolution fusion feature map F multi for:

[0095]

[0096] in, is the output feature map of the Nth node.

[0097] Specifically, in the multi-layer convolution fusion module, the above operation is repeated in each node, and the features of multiple nodes are fused to form the final feature map: multi-layer convolution fusion feature map F multi In this solution, N=6 is set.

[0098] By using a multi-layer convolutional fusion module, multiple nodes are added between the encoder and the decoder to fuse and superimpose feature information at different levels. This method effectively overcomes the problem of small target detail information loss caused by pooling operations and excessive receptive fields in traditional methods, and significantly improves the detection accuracy of the model.

[0099] Preferably, the feature map output by each node in the multi-layer convolution fusion module will be fused through a specific convolution operation in the future, and finally a multi-level feature map will be obtained, which will be input into the decoder for further processing to restore the image size and realize target detection, and finally obtain

[0100]

[0101] S103, extracting and fusing features of several receptive fields of the multi-layer convolution fusion feature map through a multi-receptive field fusion module to obtain a multi-receptive field fusion feature map;

[0102] Convolution kernels with different receptive fields can capture target features of different scales, which helps to detect small targets with a large size range. The fusion of multi-receptive field features can make up for the limitations of a single receptive field in target detection and improve the accuracy of detection.

[0103] like Figure 4 The figure shows a schematic diagram of a multi-receptive field fusion module, which uses three convolutional layers with residual learning in different receptive fields to obtain target characteristics in different local areas, and then fuses the features obtained from different receptive fields to form a multi-receptive field fusion module.

[0104] In some embodiments, the multi-receptive field fusion module extracts and fuses the features of several receptive fields of the multi-layer convolution fusion feature map to obtain the multi-receptive field fusion feature map, specifically including:

[0105] According to the convolutional layers with residual learning in three different receptive fields, the target characteristics in different local areas are obtained. The convolution kernel sizes corresponding to the three receptive fields are 3, 5, and 7 respectively. The receptive fields of different sizes are used to obtain the corresponding depth features:

[0106]

[0107]

[0108] Among them, x is the input, G i_j is the value of the corresponding convolution kernel, i is 1-3, j is 1 or 2, F i is the feature obtained after a series of operations;

[0109] Concatenate the three features according to the first dimension to obtain the multi-receptive field fusion feature map F output : F output =F1+F2+F3

[0110] Among them, F i It is the feature obtained after a series of operations, and the value of i is 1-3.

[0111] In some embodiments, a multi-receptive field fusion feature map F is obtained. output Afterwards, the method further comprises:

[0112] The multi-receptive field fusion feature map F output With the original infrared image I in Multiply the pixels and compare the multiplied result with the original infrared image I in Addition;

[0113] The added feature map is processed through the convolution layer to obtain the feature map

[0114] F in =Conv(I in +(I in ×F output ))

[0115] Among them, I in is the original infrared image, F output It is a multi-receptive field fusion feature map.

[0116] The multi-receptive field fusion module extracts features by combining receptive fields of different sizes and performs fusion processing. This technical approach enhances the model's ability to detect objects of different scales and shapes, especially in complex backgrounds, which can significantly reduce false detections and improve detection reliability.

[0117] S104, adaptively adjusting the channel weights of the multi-receptive field fusion feature map through an efficient channel attention module to obtain an adaptively adjusted feature map;

[0118] By adaptively adjusting channel weights, strengthening important features and suppressing irrelevant features, the model's attention to weak targets is improved. Compared with the traditional attention mechanism, the efficient channel attention module can maintain a low computational complexity while improving performance.

[0119] In some embodiments, the multi-receptive field fusion feature map is adaptively adjusted in channel weights through an efficient channel attention module to obtain an adaptively adjusted feature map, specifically including:

[0120] The feature map F in Perform global average pooling to obtain aggregate features:

[0121] F conv =Conv1d(F in )

[0122] The channel weights are generated by a fast 1D convolution of size k, where k is adaptively determined according to the mapping of the channel dimension C:

[0123]

[0124] Where |t| odd represents the odd number closest to t, γ and b are both 1;

[0125] The obtained feature map is normalized by SIGMOD and then fused with the multi-receptive field feature map F output Multiply to generate adaptive adjustment feature map

[0126] F CAM =σ(F conv )×F output

[0127] Among them, F output is the multi-receptive field fusion feature map, F conv is the aggregation feature.

[0128] The use of an efficient channel attention module can significantly enhance the channel weight of the feature map, thereby improving the detection effect. Compared with the existing complex attention module, this module avoids excessive complexity of the model while ensuring high performance.

[0129] Preferably, the network output is mapped to the semantic segmentation result of the original input image through the full convolution module, that is, the small target information is predicted and the detected infrared image binarization image is output.

[0130] F result =FCN(F CAM ).

[0131] S105, mapping the adaptively adjusted feature map through a fully connected layer module, obtaining the semantic segmentation result of the original infrared image and using it as the detected infrared weak target image.

[0132] The adaptively adjusted feature map is mapped to the final semantic segmentation result to achieve target detection. The fully connected layer can be adjusted according to specific task requirements, such as changing the output dimension to adapt to different numbers of target categories.

[0133] In some embodiments, after mapping the adaptively adjusted feature map through the fully connected layer module, obtaining the semantic segmentation result of the original infrared image and using it as the detected infrared dim small target image, the method further includes:

[0134] The optimal loss function is obtained based on scale-sensitive loss and positioning-sensitive loss:

[0135]

[0136] in, and They represent scale-sensitive loss and positioning-sensitive loss respectively;

[0137] The scale-sensitive loss expression is:

[0138]

[0139]

[0140] where Var(·,·) is the function that obtains the variance of the provided scalar, |A p |with|A gt The larger the gap between |, the smaller w, the greater the scale sensitivity loss, A p For a given set of predicted images, A gt is a set of truth values.

[0141] Specifically, suppose A p With Agt The IoU between them is fixed, |A p |with|A gt The larger the gap between |, the smaller w is, which leads to a large scale-sensitive loss. If the predicted and true value scales (i.e., A p and A gt The number of pixels in ) is very different, and the detector should pay more attention to the objects with larger loss.

[0142] In some embodiments, obtaining the location-sensitive loss includes:

[0143] Given a prediction image A p Sets and truth values ​​like A gt The coordinates of all pixels in the set are averaged to obtain A p and A gt The corresponding first center point and second center point are denoted as c p =(x p ,y p ) and c gt =(x gt ,y gt );

[0144] Convert the coordinates of the first center point and the second center point into the corresponding distance d in the polar coordinate system p and angle θ p :

[0145]

[0146]

[0147] Among them, d p is the corresponding distance in the polar coordinate system, θ p is the corresponding angle θ in the polar coordinate system p ;

[0148] Determine the positioning sensitive loss based on the predicted center point and the true center point of the target:

[0149]

[0150] Among them, d gt and θ gt c gt Position-sensitive loss and angular position loss in polar coordinates.

[0151] The above gt and θ gt c gtThe range and angular position losses in polar coordinates can effectively distinguish most different position errors, making the detector more sensitive to different types of position errors and thus locating the target more accurately.

[0152] In some embodiments, the intersection over union (IOU) metric is used to evaluate and verify the small target position detected by the model. First, the real target area R is defined. true and the target region R predicted by the model pred , calculate the intersection area R of the two inter =R true ∩R pred and the union region R union =R true ∩R pred Then, the IOU value is calculated as the evaluation indicator, and its formula is:

[0153]

[0154] Among them, R inter is the intersection area of ​​the two regions, R union is the union of two regions.

[0155] In summary, the infrared dim target detection model has strong feature extraction, fusion and mapping capabilities by combining modules such as multi-layer convolution fusion, multi-receptive field fusion, efficient channel attention and fully connected layer. It can improve the detection accuracy and robustness of dim targets and efficiently detect dim targets from complex infrared backgrounds.

[0156] The present application also proposes an infrared small target image detection system, the system comprising: a detection model building unit and a detection result determination unit;

[0157] The detection model building unit is used to build an infrared small target detection model. The infrared small target detection network includes a multi-layer convolution fusion module, a multi-receptive field fusion module, an efficient channel attention module and a fully connected layer module.

[0158] The detection result determination unit is used to extract and fuse the features of the original infrared image through a multi-layer convolution fusion module to obtain a multi-layer convolution fusion feature map; extract and fuse the features of several receptive fields of the multi-layer convolution fusion feature map through a multi-receptive field fusion module to obtain a multi-receptive field fusion feature map; adaptively adjust the channel weights of the multi-receptive field fusion feature map through an efficient channel attention module to obtain an adaptively adjusted feature map; map the adaptively adjusted feature map through a fully connected layer module to obtain the semantic segmentation result of the original infrared image and use it as the detected infrared weak target image.

[0159] The present application also proposes a readable storage medium storing a computer program. When the computer program is executed by a processor, the processor executes the following steps:

[0160] Construct an infrared small target detection model. The infrared small target detection network includes a multi-layer convolution fusion module, a multi-receptive field fusion module, an efficient channel attention module, and a fully connected layer module.

[0161] The original infrared image is subjected to feature extraction and fusion through a multi-layer convolution fusion module to obtain a multi-layer convolution fusion feature map;

[0162] The multi-receptive field fusion module extracts and fuses the features of several receptive fields of the multi-layer convolution fusion feature map to obtain a multi-receptive field fusion feature map;

[0163] The multi-receptive field fusion feature map is adaptively adjusted in channel weights through an efficient channel attention module to obtain an adaptively adjusted feature map;

[0164] The adaptively adjusted feature map is mapped through the fully connected layer module to obtain the semantic segmentation result of the original infrared image and use it as the detected infrared weak target image.

[0165] The present application also proposes a computer device, including a memory and a processor, wherein the memory stores a computer program, and the computer program is executed by the processor to perform the following steps:

[0166] Construct an infrared small target detection model. The infrared small target detection network includes a multi-layer convolution fusion module, a multi-receptive field fusion module, an efficient channel attention module, and a fully connected layer module.

[0167] The original infrared image is subjected to feature extraction and fusion through a multi-layer convolution fusion module to obtain a multi-layer convolution fusion feature map;

[0168] The multi-receptive field fusion module extracts and fuses the features of several receptive fields of the multi-layer convolution fusion feature map to obtain a multi-receptive field fusion feature map;

[0169] The multi-receptive field fusion feature map is adaptively adjusted in channel weights through an efficient channel attention module to obtain an adaptively adjusted feature map;

[0170] The adaptively adjusted feature map is mapped through the fully connected layer module to obtain the semantic segmentation result of the original infrared image and use it as the detected infrared weak target image.

[0171] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiments can be completed by instructing the relevant hardware through a computer program, and the program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in the present application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0172] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0173] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the patent of the present application. It should be pointed out that, for ordinary technicians in this field, several modifications and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. The above disclosure is only a preferred embodiment of the present invention, and of course it cannot be used to limit the scope of the rights of the present invention. Therefore, equivalent changes made according to the claims of the present invention are still within the scope covered by the present invention.

Claims

1. A method for detecting infrared small target images, characterized in that: The method comprises: Constructing an infrared small target detection model, wherein the infrared small target detection network includes a multi-layer convolution fusion module, a multi-receptive field fusion module, an efficient channel attention module and a fully connected layer module; The multi-layer convolution fusion module is used to extract and fuse the original infrared image to obtain a multi-layer convolution fusion feature map; The multi-receptive field fusion module extracts and fuses the features of several receptive fields of the multi-layer convolution fusion feature map to obtain a multi-receptive field fusion feature map; Adaptively adjusting the channel weights of the multi-receptive field fusion feature map through the efficient channel attention module to obtain an adaptively adjusted feature map; The adaptively adjusted feature map is mapped through the fully connected layer module to obtain the semantic segmentation result of the original infrared image and use it as the detected infrared weak target image.

2. The infrared small target image detection method according to claim 1, characterized in that: The extracting and fusing the original infrared image by the multi-layer convolution fusion module to obtain the multi-layer convolution fusion feature map specifically includes: In each convolution node of the multi-layer convolution fusion module, any original infrared image Through the convolution layer, normalization layer, activation function and residual connection operation in sequence: F conv =ReLU(BN(Conv(I in ))) Among them, the residual connection operation expression is: F out =F conv +I in ; Each convolution node is a repeated structure of convolution layer, normalization layer, activation function and residual connection. When there are N nodes, the multi-layer convolution fusion feature map F multi for: in, is the output feature map of the Nth node.

3. The infrared small target image detection method according to claim 2, characterized in that: The multi-receptive field fusion module extracts and fuses the features of several receptive fields of the multi-layer convolution fusion feature map to obtain the multi-receptive field fusion feature map, specifically including: According to the convolutional layers with residual learning in three different receptive fields, the target characteristics in different local areas are obtained, wherein the convolution kernel sizes corresponding to the three receptive fields are 3, 5, and 7 respectively, and the corresponding depth features are obtained using receptive fields of different sizes: Among them, x is the input, G i_j is the value of the corresponding convolution kernel, i is 1-3, j is 1 or 2, F i is the feature obtained after a series of operations; Concatenate the three features according to the first dimension to obtain the multi-receptive field fusion feature map F output :F output =F1+F2+F3 Among them, F i It is the feature obtained after a series of operations, and the value of i is 1-3.

4. The infrared small target image detection method according to claim 3 is characterized in that: The multi-receptive field fusion feature map F is obtained output Afterwards, the method further comprises: The multi-receptive field fusion feature map F output With the original infrared image I in Perform pixel multiplication and compare the multiplication result with the original infrared image I in Addition; The added feature map is processed through the convolution layer to obtain the feature map F in =Conv(I in +(I in ×F output )) Among them, I in is the original infrared image, F output It is a multi-receptive field fusion feature map.

5. The infrared small target image detection method according to claim 4, characterized in that: The step of adaptively adjusting the channel weights of the multi-receptive field fusion feature map by the efficient channel attention module to obtain the adaptively adjusted feature map specifically includes: The feature map F in Perform global average pooling to obtain aggregate features: F conv =Conv1d(F in ); The channel weights are generated by a fast 1D convolution of size k, where k is adaptively determined according to the mapping of the channel dimension C: Where |t| odd represents the odd number closest to t, γ and b are both 1; The obtained feature map is normalized by SIGMOD and then fused with the multi-receptive field feature map F output Multiply to generate adaptive adjustment feature map F CAM =σ(F conv )×F output Among them, F output is the multi-receptive field fusion feature map, F conv is the aggregation feature.

6. The infrared small target image detection method according to any one of claims 2 to 5, characterized in that: After mapping the adaptively adjusted feature map through the fully connected layer module, obtaining the semantic segmentation result of the original infrared image and using it as the detected infrared weak target image, the method further includes: The optimal loss function is obtained based on scale-sensitive loss and positioning-sensitive loss: in, and They represent scale-sensitive loss and positioning-sensitive loss respectively; The scale-sensitive loss expression is: where Var(·,·) is the function that obtains the variance of the provided scalar, |A p |with|A gt The larger the gap between |, the smaller w, the greater the scale sensitivity loss, A p For a given set of predicted images, A gt is a set of truth values.

7. The infrared small target image detection method according to claim 6, characterized in that: The obtaining of the positioning sensitivity loss comprises: The given prediction image A p The set and the truth value are like A gt The coordinates of all pixels in the set are averaged to obtain A p and A gt The corresponding first center point and second center point are denoted as c p =(x p ,y p ) and c gt =(x gt ,y gt ); Convert the coordinates of the first center point and the second center point into the corresponding distance d in the polar coordinate system p and angle θ p : Among them, d p is the corresponding distance in the polar coordinate system, θ p is the corresponding angle θ in the polar coordinate system p ; Determine the positioning sensitive loss based on the predicted center point and the true center point of the target: Among them, d gt and θ gt c gt Position-sensitive loss and angular position loss in polar coordinates.

8. An infrared small target image detection system, characterized in that: The system comprises: a detection model building unit and a detection result determination unit; The detection model construction unit is used to construct an infrared small target detection model, and the infrared small target detection network includes a multi-layer convolution fusion module, a multi-receptive field fusion module, an efficient channel attention module and a fully connected layer module; The detection result determination unit is used to perform feature extraction and fusion on the original infrared image through the multi-layer convolution fusion module to obtain a multi-layer convolution fusion feature map; perform feature extraction and fusion of several receptive fields on the multi-layer convolution fusion feature map through the multi-receptive field fusion module to obtain a multi-receptive field fusion feature map; perform adaptive adjustment of channel weights on the multi-receptive field fusion feature map through the efficient channel attention module to obtain an adaptively adjusted feature map; map the adaptively adjusted feature map through the fully connected layer module to obtain a semantic segmentation result of the original infrared image and use it as a detected infrared weak target image.

9. A readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor is caused to execute the steps of the method according to any one of claims 1 to 7.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the method according to any one of claims 1 to 7.

Images