Infrared image target detection network

Abstract

This application provides an infrared image target detection network, comprising: a Focus module whose input is an infrared image; an MP_CBL module including two branches; the first branch is downsampled by a MaxPool module (2×2 size, stride 2, no padding), and then transformed by a CBL module; the second branch first extracts features through a CBL module while halving its width and height, and then transforms the channel number through another CBL module; the feature maps processed by the first and second branches are then stitched together by a stitching module. The detection neck includes a feature pyramid network structure and a path aggregation network structure, used to process the features extracted from the network backbone; the detection head is used for target classification and localization prediction. This application improves the target detection accuracy of the network, enabling it to identify targets in harsh environments, achieving network lightweighting, and allowing deployment on edge devices.

Description

Technical Field

[0001] This application relates to the field of infrared image target detection, specifically to an infrared image target detection network. Background Technology

[0002] In the traditional field of object detection, most object detection networks are trained and inferred based on images acquired by visible light cameras. However, these methods perform poorly in harsh environments such as sandstorms, smog, and low light conditions. This is mainly because visible light imaging relies on external light sources and is easily affected by ambient light and particle scattering, leading to blurred or even lost target features, making it difficult to meet the requirements of all-weather, highly robust monitoring. In addition, existing object detection networks (such as Faster R-CNN and YOLO) typically have high computational complexity and a large number of parameters, making them difficult to deploy on edge devices.

[0003] In contrast, infrared cameras, based on the thermal radiation imaging of objects, do not rely on external light sources. Their wavelengths are on the order of micrometers, effectively penetrating harsh weather conditions such as smoke and haze, reducing scattering interference, and thus providing more stable target feature information. However, infrared images are characterized by low contrast, limited texture information, and low signal-to-noise ratio, making it difficult for existing target detection networks to accurately extract effective features. Furthermore, existing target detection networks are not optimized for the characteristics of infrared imaging, resulting in computational redundancy and failing to meet the lightweight requirements of edge devices (e.g., security equipment).

[0004] Therefore, there is an urgent need for a lightweight target detection network for infrared images that can operate stably in harsh environments and be deployed on edge devices to meet the needs of all-weather intelligent monitoring. Summary of the Invention

[0005] In view of the above problems, this application provides an infrared image target detection network that overcomes or at least partially solves the problems that existing target detection networks cannot operate stably in harsh environments and cannot be deployed on edge devices.

[0006] A first aspect of this application provides an infrared image target detection network, including: a network backbone, a neck detection module, and a head detection module. The network backbone includes a Focus module, a CBL module, a first CSP module, an Attention module, an MP_CBL module, a second CSP module, and an SPPF module, wherein the Attention module is an attention mechanism module. The input to the Focus module is an infrared image. The Focus module, the first CBL module, the second CBL module, the first CSP module, the first Attention module, the first MP_CBL module, the second CSP module, the second Attention module, the second MP_CBL module, the third CSP module, the third Attention module, the third MP_CBL module, and the SPPF module are connected sequentially. The first CBL module and the second CBL module are CBL modules of the same type; the second CSP module and the third CSP module are CSP modules of the same type; the first Attention module, the second Attention module, and the third Attention module are Attention modules of the same type; and the first MP_CBL module, the second MP_CBL module, and the third MP_CBL module are MP_CBL modules of the same type.

[0007] The MP_CBL module consists of two branches. The first branch performs downsampling through the MaxPool module (2×2 size, stride 2, no padding) and then channel transformation through the CBL module. The second branch first extracts features through the CBL module while halving the width and height, and then performs channel transformation through another CBL module. The feature maps processed by the first and second branches are then concatenated using a stitching module.

[0008] The detection neck consists of a feature pyramid network structure and a path aggregation network structure. The detection neck processes the features extracted from the network backbone and outputs the processed feature map to the detection head. The detection head includes convolutional modules and a detection head module for object classification and localization prediction.

[0009] In this embodiment, the Focus module segments the input infrared image into multi-dimensional data, quadrupling the input channels to distribute feature information across the channel space. A convolution operation is then performed to obtain a downsampled feature map, reducing the computational load of the infrared image target detection network while minimizing information loss due to downsampling, thus maintaining the network's detection accuracy. The CBL module initially extracts low-level features (such as edges and textures), the first CSP module reduces computational redundancy and enhances gradient flow, and the Attention module, while ensuring the network's lightweight design, allows it to learn the importance of different channels, improving computational efficiency and accuracy. The MP_CBL module combines max pooling and convolution operations, without introducing parameters to the network, reducing the feature map size while retaining key feature information to balance speed and accuracy. The second CSP module enhances feature extraction capabilities. The SPPF module allows the infrared image target detection network to perform max pooling on the input feature map at different scales. Finally, these pooled features are concatenated to form a fixed-size feature vector, which helps improve the accuracy and robustness of the infrared image target detection network for target detection, and provides better adaptability to targets at different scales. The detection neckline processes the features extracted from the network backbone and outputs the processed feature map to the detection head. The detection head is used for target classification and localization prediction.

[0010] Thus, the infrared image target detection network provided in this application improves upon existing target detection networks by adding Attention and MP_CBL modules, thereby enhancing the network's target detection accuracy and enabling it to be used to detect infrared images for target identification in harsh environments. Furthermore, because the infrared image target detection network provided in this application reduces the number of network parameters through the MP_CBL module, it achieves network lightweighting and can therefore be deployed on edge devices.

[0011] In one alternative approach, the Attention module includes a GAP module, a convolutional module, an average pooling module, a concatenation module, a batch normalization module, a split module, a sigmoid module, and a reweight module. The GAP module is connected to the convolutional module; the outputs of the convolutional modules are each connected to an average pooling module; the outputs of the two average pooling modules are connected to the concatenation module; the convolutional module is connected to the output of the concatenation module; the batch normalization module is connected to the output of the convolutional module; the split module is connected to the output of the batch normalization module; the output of the split module is sequentially connected to a convolutional module and a sigmoid module; and the outputs of the two sigmoid modules are connected to the reweight module.

[0012] In this embodiment, the Attention module uses the GAP module to perform global average pooling on the input feature map to generate channel statistical features. After passing through the convolution module, the channel statistical features are split into horizontal and vertical branches, which are then processed by the average pooling module. Next, the two-branch features are fused by the concatenation module, and feature transformation is performed by the convolution module and the batch normalization module. Then, the features are split again by the Split module, and the split features are convolved. Attention weights are generated by the Sigmoid activation function in the Sigmoid module. Finally, the attention weights are multiplied with the original input features by the Reweight module to output the optimized feature map. The above process achieves the coordinated optimization of channel and spatial information through a multi-branch split-fusion mechanism.

[0013] In one alternative approach, neck detection includes a third CSP module, an Attention module, and a CBL module, with an Attention module set between each third CSP module and each CBL module.

[0014] In this embodiment, the infrared image target detection network provided by this application adopts the feature pyramid network structure and path aggregation network structure commonly found in existing target recognition networks. The main improvement lies in setting an Attention module between each third CSP module and CBL module. The neck detection uses the third CSP module to extract basic features, which are then enhanced in a recurrent unit composed of the Attention module and the CBL module, with upsampling used alternately to expand resolution. The entire process reduces computational redundancy through the third CSP module and dynamically optimizes channel and spatial attention using the Attention module, achieving efficient multi-scale information flow and fusion.

[0015] In one alternative approach, the detection head includes a multi-scale convolutional module and a detection head module.

[0016] In one alternative approach, the Focus module includes multiple convolutional modules with different kernel sizes and a concatenation module, with the outputs of the multiple convolutional modules with different kernel sizes connected to the concatenation module.

[0017] In one alternative approach, the CBL module includes a convolution module, a batch normalization module, and a leakage correction linear activation module, which are connected sequentially.

[0018] In one alternative embodiment, the first CSP module includes: a CBL module, a residual component, a convolution module, and a stitching module. The first CSP module includes a third branch and a fourth branch. Within the third branch, the CBL module, residual component, convolution module, and stitching module are sequentially connected. Within the fourth branch, the input of the CBL module is connected to the input of the first CSP module, and the output of the CBL module is connected to the stitching module.

[0019] In one alternative embodiment, the SPPF module includes a CBL module, a first max pooling module, a second max pooling module, a third max pooling module, and a splicing module. The output of the CBL module is connected to the first max pooling module, the second max pooling module, and the third max pooling module, respectively, and the outputs of the first max pooling module, the second max pooling module, and the third max pooling module are connected to the splicing module.

[0020] In one alternative embodiment, the second CSP module includes: a CBL module, multiple residual components (e.g., three), a convolution module, and a stitching module. The second CSP module includes a fifth branch and a sixth branch. In the fifth branch, the CBL module, multiple residual components, the convolution module, and the stitching module are sequentially connected. In the sixth branch, the input of the CBL module is connected to the input of the second CSP module, and the output of the CBL module is connected to the stitching module.

[0021] In one alternative embodiment, the third CSP module includes a CBL module, a convolution module, and a stitching module. The third CSP module includes a seventh branch and an eighth branch. In the seventh branch, the CBL module, the convolution module, and the stitching module are connected sequentially. In the eighth branch, the input of the CBL module is connected to the input of the third CSP module, and the output of the CBL module is connected to the stitching module.

[0022] The above description is merely an overview of the technical solutions of the embodiments of this application. In order to better understand the technical means of the embodiments of this application and to implement them in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the embodiments of this application more obvious and understandable, specific implementation methods of this application are described below. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a schematic diagram of an infrared image target detection network structure provided for some embodiments of this application.

[0025] Figure 2 The diagram shows the structure of the MP_CBL module provided in some embodiments of this application.

[0026] Figure 3 This is a schematic diagram of the structure of the Attention module provided in some embodiments of this application.

[0027] Figure 4 This is a schematic diagram of the structure of the Focus module provided in some embodiments of this application.

[0028] Figure 5 This is a schematic diagram of the structure of the CBL module provided in some embodiments of this application.

[0029] Figure 6 This is a schematic diagram of the structure of a first CSP module provided in some embodiments of this application.

[0030] Figure 7 This is a schematic diagram of the structure of an SPPF module provided in some embodiments of this application.

[0031] Figure 8 This is a schematic diagram of the structure of a third CSP module provided in some embodiments of this application.

[0032] Figure 9 The image shows the detection effect of an infrared image target detection network provided in some embodiments of this application. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0034] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs; the terminology used herein in the specification of the application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application.

[0035] The terms "comprising" and "having," and any variations thereof, used in the specification, claims, and drawings of this application are intended to cover without excluding other meanings. The words "a" or "an" do not exclude the presence of multiples.

[0036] The term "embodiment" as used herein means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of the phrase "embodiment" in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0037] Furthermore, the terms "first," "second," etc., in the specification and claims of this application or in the aforementioned drawings are used to distinguish different objects rather than to describe a specific order, and may explicitly or implicitly include one or more of the features.

[0038] In the description of this application, unless otherwise stated, "multiple" means two or more (including two), and similarly, "multiple groups" means two or more (including two groups).

[0039] In the description of this application, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linkage" should be interpreted broadly. For example, "connection" or "linkage" in mechanical structures can refer to a physical connection, such as a fixed connection, for example, a connection fixed by fasteners, such as a connection fixed by screws, bolts, or other fasteners; a physical connection can also be a detachable connection, such as a snap-fit ​​or interlocking connection; a physical connection can also be an integral connection, such as a connection formed by welding, bonding, or integral molding. In circuit structures, "connection" or "linkage" can refer not only to a physical connection but also to an electrical connection or a signal connection. For example, it can be a direct connection, i.e., a physical connection, or an indirect connection through at least one intermediate component, as long as the circuit is connected; it can also refer to the internal connection of two components. Signal connection can refer not only to signal connection through a circuit but also to signal connection through a media, such as radio waves. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0040] Figure 1 This is a schematic diagram of an infrared image target detection network structure provided in some embodiments of this application. Figure 2 This is a schematic diagram illustrating the structure of the MP_CBL module provided in some embodiments of this application. (Reference) Figure 1 and Figure 2The infrared image target detection network provided in this application includes: a network backbone, a neck detection module, and a head detection module. The network backbone includes a Focus module, a CBL module, a first CSP module (CSP1_1 in the figure), an Attention module, an MP_CBL module, a second CSP module (CSP1_3 in the figure), and an SPPF module, wherein the Attention module is an attention mechanism module. The input to the Focus module is an infrared image. The Focus module, the first CBL module, the second CBL module, the first CSP module, the first Attention module, the first MP_CBL module, the second CSP module, the second Attention module, the second MP_CBL module, the third CSP module, the third Attention module, the third MP_CBL module, and the SPPF module are connected sequentially. The first CBL module and the second CBL module are the same type of CBL module (hereinafter referred to as CBL module). Figure 1 (labeled as CBL module), the second CSP module and the third CSP module are the same type of CSP module (both are CSP_3, Figure 1 (CSP_3 is used as the label in the text). The first Attention module, the second Attention module, and the third Attention module are all Attention modules of the same type (hereinafter referred to as Attention modules). Figure 1 (labeled as Attention modules in the text), the first MP_CBL module, the second MP_CBL module, and the third MP_CBL module are all MP_CBL modules of the same type (hereinafter referred to as MP_CBL modules). Figure 1 (The module is labeled MP_CBL).

[0041] The MP_CBL module consists of two branches. The first branch performs downsampling through the MaxPool module (2×2 size, stride 2, no padding) and then channel transformation through the CBL module. The second branch first performs convolution operation through a CBL module (3×3 size, stride 2, padding 1) to extract features while halving the width and height, and then performs channel transformation through another CBL module. The feature maps processed by the first branch and the feature maps processed by the second branch are then concatenated through the Concate module.

[0042] The detection neck consists of a feature pyramid network structure and a path aggregation network structure. The detection neck processes the features extracted from the network backbone and outputs the processed feature map to the detection head. The detection head includes a convolutional module (Conv) and a detection module (Detect) for object classification and localization prediction.

[0043] In this embodiment, the Focus module segments the input infrared image into multi-dimensional data, quadrupling the input channels to distribute feature information across the channel space. A convolution operation is then performed to obtain a downsampled feature map, reducing the computational load of the infrared image target detection network while minimizing information loss due to downsampling, thus maintaining the network's detection accuracy. The CBL module initially extracts low-level features (such as edges and textures), the first CSP module reduces computational redundancy and enhances gradient flow, and the Attention module, while ensuring the network's lightweight design, allows it to learn the importance of different channels, improving computational efficiency and accuracy. The MP_CBL module combines max pooling and convolution operations, without introducing parameters to the network, reducing the feature map size while retaining key feature information to balance speed and accuracy. The second CSP module enhances feature extraction capabilities. The SPPF module allows the infrared image target detection network to perform max pooling on the input feature map at different scales. Finally, these pooled features are concatenated to form a fixed-size feature vector, which helps improve the accuracy and robustness of the infrared image target detection network for target detection, and provides better adaptability to targets at different scales. The detection neckline processes the features extracted from the network backbone and outputs the processed feature map to the detection head. The detection head is used for target classification and localization prediction.

[0044] Thus, the infrared image target detection network provided in this application improves upon existing target detection networks by adding Attention and MP_CBL modules, thereby enhancing the network's target detection accuracy and enabling it to be used to detect infrared images for target identification in harsh environments. Furthermore, because the infrared image target detection network provided in this application reduces the number of network parameters through the MP_CBL module, it achieves network lightweighting and can therefore be deployed on edge devices.

[0045] In some embodiments, Figure 3 This is a schematic diagram illustrating the structure of an Attention module provided in some embodiments of this application. (Reference) Figure 3The Attention module includes a GAP module, a Convolutional module (Conv), an Average Pooling module (AveragePool), a Concatenating module (Concate), a Batch Normalization (BN) module, a Split module, a Sigmoid module, and a Reweight module. The GAP module is connected to the Convolutional module (Conv). The outputs of the Convolutional module (Conv) are each connected to one Average Pooling module (AveragePool). The outputs of both Average Pooling modules (AveragePool) are connected to the Concatenating module (Concate). The Convolutional module (Conv) is connected to the output of the Concatenating module (Concate). The Batch Normalization (BN) module is connected to the output of the Convolutional module (Conv). The Split module is connected to the output of the Batch Normalization (BN) module. The outputs of the Split module are sequentially connected to one Convolutional module (Conv) and one Sigmoid module (Sigmoid). The outputs of both Sigmoid modules are connected to the Reweight module.

[0046] In practical applications, the GAP module is connected to the Conv module. The output of the Conv module can also be connected to the Batch Normalization module BN. The output of the Batch Normalization module BN is then connected to an Average Pooling module AveragePool.

[0047] In this embodiment, the Attention module uses the GAP module to perform global average pooling on the input feature map to generate channel statistical features, which are then processed by the Conv module. This enhances the network's modeling ability in the channel dimension, improves computational efficiency, and does not increase the number of network parameters, thus ensuring the network's lightweight nature. The channel statistical features are then decomposed into horizontal P... h and vertical P w The two branches are processed by the AveragePool module; then, the features from both branches are fused by the Concate module, followed by feature transformation through the Convolution module (Conv) and the Batch Normalization (BN) module; next, the features are split by the Split module, and the split features are convolved, with attention weights generated by the Sigmoid activation function in the Sigmoid module; finally, the attention weights are multiplied by the original input features through the Reweight module to output the optimized feature map. In this way, different attention weights can be assigned to different coordinates of the target's bounding box. By learning the attention weights, the degree of attention to different coordinates is adaptively adjusted, thereby improving the prediction accuracy of coordinate positions and ultimately improving the network's detection accuracy. The above process, through a multi-branch split-fusion mechanism, achieves the coordinated optimization of channel and spatial information.

[0048] In some embodiments, the neck detection includes a third CSP module, an Attention module, and a CBL module, with an Attention module set between each third CSP module and CBL module.

[0049] In this embodiment, the infrared image target detection network provided in this application adopts the feature pyramid network structure and path aggregation network structure commonly found in existing target recognition networks. See [link / reference] for details. Figure 1 This application does not elaborate on the detailed structure of the neck detection mechanism. Its main improvement lies in setting an Attention module between each third CSP module and CBL module. Neck detection extracts basic features through the third CSP module, then enters a recurrent unit composed of the Attention and CBL modules for feature enhancement, and alternately uses upsampling to expand resolution. The entire process reduces computational redundancy through the third CSP module and dynamically optimizes channel and spatial attention using the Attention module, achieving efficient multi-scale information flow and fusion.

[0050] In some embodiments, the detection head includes a multi-scale convolutional module Conv and a detection head module Detect.

[0051] refer to Figure 1 The detection head consists of three convolutional modules (Conv) and a detection head module (Detect), meaning the detection head includes three convolutional modules (Conv) and a detection head module (Detect) of different scales.

[0052] The input to each convolutional module (Conv) comes from different stages of the network backbone, corresponding to high-level, mid-level, and low-level features at different resolutions. High-level features are used to detect large objects, mid-level features to detect medium-sized objects, and low-level features to detect small objects. Thus, the infrared image target detection network provided in this application can better adapt to infrared images of different resolutions, achieving efficient and accurate target localization and classification.

[0053] In some embodiments, Figure 4 This is a schematic diagram illustrating the structure of a Focus module provided in some embodiments of this application. (Reference) Figure 4 The Focus module includes multiple convolutional modules (Conv) with different kernel sizes and a concatenation module (Concate). For example, it can include the outputs of four convolutional modules (Conv) with different kernel sizes (1x1, 3x3, 5x5, and 7x7) connected to the concatenation module (Concate). In this way, convolutions with different kernel sizes can be used to extract features at multiple scales, improving the network's performance at different resolutions.

[0054] In some embodiments, Figure 5 This is a schematic diagram of the structure of a CBL module provided in some embodiments of this application. (Reference) Figure 5 The CBL module includes a convolution module (Conv), a batch normalization module (BN), and a leakage correction linear activation module (LeakyReLU), which are connected sequentially.

[0055] In some embodiments, Figure 6 This is a schematic diagram of the structure of a first CSP module provided in some embodiments of this application. In this embodiment, the first CSP module is CSP1_1, therefore, referring to... Figure 6 The diagram includes a residual component, ResUnit. The first CSP module includes: a CBL module, a residual component, ResUnit, a convolution module, Conv, and a concatenation module, Concate. The first CSP module includes a third branch and a fourth branch. In the third branch, the CBL module, ResUnit, Conv, and Concate are connected sequentially. In the fourth branch, the input of the CBL module is connected to the input of the first CSP module, and the output of the CBL module is connected to the Concate module.

[0056] In practical applications, the second CSP module is CSP1_3; therefore, refer to... Figure 6 The diagram includes three residual components, ResUnit. The second CSP module comprises a CBL module, a residual component ResUnit, a convolution module Conv, and a concatenation module Concate. The second CSP module has a fifth branch and a sixth branch. In the fifth branch, the CBL module, the residual component ResUnit, the convolution module Conv, and the concatenation module Concate are connected sequentially. In the sixth branch, the input of the CBL module is connected to the input of the second CSP module, and the output of the CBL module is connected to the concatenation module Concate.

[0057] In this embodiment, basic feature extraction is performed using the CBL module, deep feature learning is performed using the ResUnit residual component, and the Conv convolution module adjusts the channel dimensions. The processed features are then fused with the original input features using the Concate concatenation module. This process achieves gradient splitting through cross-stage partial connections and avoids network degradation by combining the ResUnit residual component design, thus realizing efficient feature extraction, gradient optimization, and network lightweighting.

[0058] In some embodiments, Figure 7 This is a schematic diagram illustrating the structure of an SPPF module provided in some embodiments of this application. (Reference) Figure 7The SPPF module includes a CBL module, a first max-pooling module (2x2MaxPool), a second max-pooling module (4x4MaxPool), a third max-pooling module (8x8MaxPool), and a concatenation module (Concate). The output of the CBL module is connected to the first max-pooling module (2x2MaxPool), the second max-pooling module (4x4MaxPool), and the third max-pooling module (8x8MaxPool), respectively. The outputs of the first max-pooling module (2x2MaxPool), the second max-pooling module (4x4MaxPool), and the third max-pooling module (8x8MaxPool) are connected to the concatenation module (Concate). Thus, by setting different max-pooling modules, the parallelism of the processing can be increased, thereby improving the computational efficiency of the network.

[0059] In some embodiments, Figure 8 This is a schematic diagram of the structure of a third CSP module provided in some embodiments of this application. In this embodiment, the third CSP module is CSP2_1; therefore, referring to... Figure 8 The diagram includes a CBL module. (Reference) Figure 8 The third CSP module includes: a CBL module, a convolution module (Conv), and a concatenation module (Concate). The third CSP module has a seventh branch and an eighth branch. In the seventh branch, the CBL module, the convolution module (Conv), and the concatenation module (Concate) are connected sequentially. In the eighth branch, the input of the CBL module is connected to the input of the third CSP module, and the output of the CBL module is connected to the concatenation module (Concate).

[0060] In neural network algorithm research, understanding a network can be achieved by deleting certain modules and studying the changes in network performance. For example, studying changes in data such as the detection accuracy, speed, and model parameter size of an object detection network is called an ablation experiment. This embodiment modifies the Attention and MP_CBL modules of the infrared image target detection network (IR) provided in this application and compares the experimental data with the YOLO algorithm network, as shown in Table 1.1.

[0061] Table 1.1 Comparison of IR ablation experimental data

[0062]

[0063] Comparing the experimental data from Experiments 1 and 2, the proposed IR significantly improves mAP compared to YOLO, reduces the number of parameters to 6.97M, and decreases the complexity metric (Giga Floating-point Operations Per Second, GFLOPs) by 0.3. This demonstrates that the proposed IR achieves significant progress in lightweight network optimization for edge devices. Comparing Experiments 1 and 4, the addition of the MP_CBL module significantly reduces the number of parameters while improving recognition accuracy and reducing model complexity. Comparing Experiments 2 and 4, the removal of the Attention module leads to a decrease in average detection accuracy (mAP), while the increase in the number of parameters and complexity metric is not significant. Therefore, the use of the Attention module can improve detection accuracy with a relatively small increase in the number of parameters.

[0064] Figure 9 The image shows the detection effect of an infrared image target detection network provided in some embodiments of this application. (Reference) Figure 9 It is evident that the infrared image target detection network provided in this application performs well in identifying targets under harsh environments.

[0065] Those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments, combinations of features from different embodiments are intended to be within the scope of this application and form different embodiments. For example, in the claims, any of the claimed embodiments can be used in any combination.

[0066] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. An infrared image target detection network, characterized in that, The network includes: a network backbone, a detection neck, and a detection head; The network backbone includes a Focus module, a CBL module, a first CSP module, an Attention module, an MP_CBL module, a second CSP module, and an SPPF module, wherein the Attention module is an attention mechanism module; The input to the Focus module is an infrared image. The Focus module, the first CBL module, the second CBL module, the first CSP module, the first Attention module, the first MP_CBL module, the second CSP module, the second Attention module, the second MP_CBL module, the third CSP module, the third Attention module, the third MP_CBL module, and the SPPF module are connected in sequence. The first CBL module and the second CBL module are CBL modules of the same type, the second CSP module and the third CSP module are CSP modules of the same type, the first Attention module, the second Attention module, and the third Attention module are Attention modules of the same type, and the first MP_CBL module, the second MP_CBL module, and the third MP_CBL module are MP_CBL modules of the same type. The MP_CBL module includes two branches. The first branch is downsampled by the MaxPool module, which has a size of 2×2, a stride of 2, and no padding, and then undergoes channel number transformation by the CBL module. The second branch first extracts features by the CBL module while halving the width and height, and then undergoes channel number transformation by another CBL module. The feature map processed by the first branch and the feature map processed by the second branch are then stitched together by the stitching module. The detection neck includes a feature pyramid network structure and a path aggregation network structure. The detection neck is used to process the features extracted by the network backbone and output the feature map processed by the detection neck to the detection head. The detection head includes a convolution module and a detection head module for target classification and localization prediction.

2. The network according to claim 1, characterized in that, The Attention module includes a GAP module, a convolution module, an average pooling module, a splicing module, a batch normalization module, a Split module, a Sigmoid module, and a Reweight module; The GAP module is connected to the convolution module. The output of the convolution module is connected to one of the average pooling modules. The outputs of the two average pooling modules are connected to the splicing module. The output of the convolution module is connected to the splicing module. The batch normalization module is connected to the output of the convolution module. The Split module is connected to the output of the batch normalization module. The output of the Split module is connected to one of the convolution modules and one of the sigmoid modules in sequence. The outputs of the two sigmoid modules are connected to the Reweight module.

3. The network according to claim 2, characterized in that, The detection neck includes a third CSP module, the Attention module, and the CBL module, with an Attention module set between each of the third CSP module and the CBL module.

4. The network according to claim 1, characterized in that, The detection head includes a multi-scale convolution module and a detection head module.

5. The network according to claim 1, characterized in that, The Focus module includes multiple convolutional modules with different kernel sizes and a splicing module, and the outputs of the multiple convolutional modules with different kernel sizes are connected to the splicing module.

6. The network according to claim 1, characterized in that, The CBL module includes a convolution module, a batch normalization module, and a leakage correction linear activation module, which are connected in sequence.

7. The network according to claim 1, characterized in that, The first CSP module includes: the CBL module, the residual component, the convolution module, and the stitching module; The first CSP module includes a third branch and a fourth branch, in which the CBL module, the residual component, the convolution module and the stitching module are connected in sequence; In the fourth branch, the input terminal of the CBL module is connected to the input terminal of the first CSP module, and the output terminal of the CBL module is connected to the splicing module.

8. The network according to claim 1, characterized in that, The SPPF module includes a CBL module, a first max pooling module, a second max pooling module, a third max pooling module, and a splicing module; The output of the CBL module is connected to the first max pooling module, the second max pooling module, and the third max pooling module, respectively. The outputs of the first max pooling module, the second max pooling module, and the third max pooling module are connected to the splicing module.

9. The network according to claim 1, characterized in that, The second CSP module includes: the CBL module, multiple residual components, the convolution module, and the stitching module; The second CSP module includes a fifth branch and a sixth branch, in which the CBL module, the plurality of residual components, the convolution module and the stitching module are connected in sequence; In the sixth branch, the input terminal of the CBL module is connected to the input terminal of the second CSP module, and the output terminal of the CBL module is connected to the splicing module.

10. The network according to claim 3, characterized in that, The third CSP module includes: the CBL module, the convolution module, and the splicing module; The third CSP module includes a seventh branch and an eighth branch, in which the CBL module, the convolution module, and the splicing module are connected in sequence; In the eighth branch, the input terminal of the CBL module is connected to the input terminal of the third CSP module, and the output terminal of the CBL module is connected to the splicing module.

Images