Visible light and infrared light fused unmanned aerial vehicle image detection method and system

Abstract

The invention discloses an unmanned aerial vehicle image detection method and system fusing visible light and infrared light, and relates to a target detection technology, and the method comprises the steps: starting a visible light sensor and an infrared sensor in groups, and carrying out the unmanned aerial vehicle detection of a target region according to a preset detection strategy; feature extraction is performed on the detected infrared image and visible light image by using independent branches, and a Res-CBAM attention mechanism module is arranged between the two independent branches; carrying out feature fusion on the image features extracted by the two branches by adopting a BiFPN structure; and based on the fused image features, performing target detection based on the classification network and the regression network to output a classification result. The target feature information in the infrared features and the scene detail information in the visible light features are fused, the complementarity of the features is enhanced, and the comprehensive performance of the model is improved.

Description

Technical Field

[0001] The present application relates to the field of target detection technology, and in particular to a method and system for detecting unmanned aerial vehicle images by integrating visible light and infrared. Background Art

[0002] With the rapid development of drone technology, its application in military reconnaissance, geographic mapping, disaster monitoring and other fields is becoming more and more extensive. However, the security issues that come with it cannot be ignored. For example, unauthorized drones may violate privacy, interfere with civil aviation flight safety, and even be used for illegal activities. In addition, the popularity of drones has also raised new challenges to airspace management and supervision. How to effectively identify and track potential threats has become a pressing issue. Therefore, it is particularly important to develop efficient drone detection technology.

[0003] In the past, drone image detection systems have mainly relied on single visible light or infrared imaging technology. While visible light imaging technology can provide high resolution and rich color information, it does not perform well at night or in low light conditions. In contrast, infrared imaging technology is able to maintain performance under these conditions, but it cannot capture the same level of detail and color information as visible light. This limitation means that these systems may not be able to accurately detect and identify targets in certain environments, especially when the target and background do not have obvious differences in thermal radiation. In addition, past methods often operate independently when processing visible light and infrared image data, lacking an effective feature fusion mechanism, resulting in the loss of important information, which in turn affects the accuracy and real-time performance of detection. In the absence of an effective fusion mechanism, the system may not be able to fully utilize the complementary information in the two image data, thereby limiting the detection performance. Summary of the invention

[0004] The embodiments of the present application provide a method and system for drone image detection that integrates visible light and infrared, which integrates target feature information in infrared features with scene detail information in visible light features, enhances feature complementarity and improves the overall performance of the model.

[0005] The present application embodiment provides a method for detecting drone images by integrating visible light and infrared, including:

[0006] Start visible light sensors and infrared sensors in groups and perform drone detection in the target area according to the preset detection strategy;

[0007] For the detected infrared image and visible light image, independent branches are used for feature extraction respectively, and a Res-CBAM attention mechanism module is set between the two independent branches;

[0008] The image features extracted by the two branches are fused using the BiFPN structure;

[0009] Based on the fused image features, target detection is performed based on the classification network and regression network to output the classification results.

[0010] Optionally, the visible light sensor and the infrared sensor are activated in groups, and drone detection is performed on the target area according to a preset detection strategy, including:

[0011] After initialization, the visible light sensor and the infrared sensor are activated, and scanning is performed according to the set radius, scanning adjustment interval, and scanning angle.

[0012] Optionally, for the detected infrared image and the visible light image, using independent branches to perform feature extraction respectively includes:

[0013] In the visible light branch for extracting image features of the visible light image, the visible light image is input into a stem sampling layer (Stem Layer), the Stem Layer includes three convolutional networks with the same structure, one of which consists of a convolutional layer with a convolution kernel size of 3×3 and a stride of 2, a batch normalization layer and a SiLU function, and the other two convolutional networks consist of a convolutional layer with a convolution kernel size of 3×3 and a stride of 1, a batch normalization layer and a SiLU function. The feature map processed by the Stem Layer is output through four CSPResLayers with a stride of 2 and the number of channels doubled successively to output a feature map with a set number of channels;

[0014] In the infrared image branch for extracting image features of the infrared image, the infrared image is input into a network with the same structure as that for extracting image features of the visible light image to output a feature map with a set number of channels.

[0015] Optionally, the Res-CBAM attention mechanism module set between two independent branches includes:

[0016] The infrared image features and visible light image features input in groups are concatenated to form the fusion feature F in ;

[0017] The fusion feature F in After a channel attention module is processed to obtain the output features The channel attention module includes a global maximum pooling layer, a global average pooling layer and a shared multi-layer perceptron;

[0018] The fusion feature F in and output features Perform element-wise multiplication to obtain the first output feature

[0019] The output features After a spatial attention module, the second output feature is obtained The spatial attention module includes a global maximum pooling layer, a global average pooling layer, two convolutional layers with a convolution kernel size of 3×3 and a Sigmoid function;

[0020] For the first output feature and the second output feature Perform element-wise multiplication to obtain the third output feature

[0021] The fusion feature F in and the third output feature Residual connection output to obtain feature map F out .

[0022] Optionally, the image features extracted by the two branches are fused using a BiFPN structure, including:

[0023] In the Neck part, the fusion feature F of some infrared image features and visible light image features is f1 、F f2 and F f3 Further fusion is performed through the BiFPN structure, and a Res-CBAM attention mechanism module is set at the end of the BiFPN structure to process the features;

[0024] The fusion feature F f1 、F f2 and F f3 In the first BiFPBN Block, a top-down network is used for processing, starting from the input fusion feature F f3 After being processed by the first CSPResLayer, the features are obtained

[0025] The characteristics Pass a convolution layer with a convolution kernel size of 1×1 and upsample the features;

[0026] The doubled-size features obtained by upsampling are combined with the fusion features F f2 Perform horizontal splicing operation, and then further process it through the second CSPResLayer to obtain the feature

[0027] The characteristics Through a 1×1 convolution layer and upsampling to the feature size and fusion feature F f1 Same as above, and the processed features are combined with the fusion features F f1 Perform horizontal splicing operation and obtain features through the third CSPResLayer processing

[0028] The characteristics After a convolution layer with a kernel size of 3×3 and a stride of 2, the features with half the size are obtained and combined with the features Perform splicing operations;

[0029] The concatenated features are processed by the fourth CSPResLayer to obtain features The characteristics And the fusion feature F f2 Add and fuse to get features

[0030] The characteristics After a convolution layer with a kernel size of 3×3 and a stride of 2, the features with half the size are obtained and combined with Perform the splicing operation and process the spliced ​​features through the fifth CSPResLayer to obtain features The characteristics And the fusion feature F f3 Add and fuse to get features

[0031] The characteristics feature feature As the input of the next BiFPN Block.

[0032] Optionally, the image features extracted by the two branches are fused using a BiFPN structure, including:

[0033] In the Neck part, three BiFPN Blocks are used continuously to obtain the output fusion feature. and as well as,

[0034] The output features of the three scales are input into three corresponding Res-CBAM attention mechanism modules respectively to output the output features of the Neck part. and

[0035] Optionally, based on the fused image features, target detection based on the classification network and the regression network includes:

[0036] In the Head part, a decoupled detection head is used to input the classification network and regression network respectively to independently complete the classification task and regression task, where:

[0037] In the classification network, the input fusion features are passed through the first convolutional network, which contains two 2D convolutional layers with 3×3 convolution kernels, a 2D batch normalization layer and a SiLU function, and then continue to pass through the second convolutional network, which contains a 2D convolutional layer with 3×3 convolution kernels, a 2D batch normalization layer and a SiLU function. Then, the features are passed through a 2D convolutional layer with a 1×1 convolution kernel, and the final classification results are output. The Distribution Focal Loss and CIoU Loss loss functions are used for training in the classification network.

[0038] In the regression network, the input fusion features are processed using the same structure as the classification network, and the BCE Loss function is used for training in the regression network to obtain the regression results.

[0039] An embodiment of the present application also proposes a drone image detection system that integrates visible light and infrared, including a processor and a memory, wherein a computer program is stored on the memory, and when the computer program is executed by the processor, the steps of the drone image detection method that integrates visible light and infrared are implemented as described above.

[0040] The embodiment of the present application fuses the target feature information in the infrared feature with the scene detail information in the visible light feature to enhance the complementarity of the features and improve the overall performance of the model. By focusing on key features and suppressing irrelevant information, the attention mechanism not only enhances the efficiency and interpretability of the detection system, but also improves the adaptability and flexibility of the model to environmental changes and drone features under different conditions.

[0041] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present application. Also, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0043] Figure 1 The following is a schematic diagram of the basic process of the drone image detection method according to an embodiment of the present application;

[0044] Figure 2 The overall network architecture of the drone image detection method according to the embodiment of the present application is shown;

[0045] Figure 3 This is a schematic diagram of a two-stream backbone feature extraction network architecture of a drone image detection method according to an embodiment of the present application;

[0046] Figure 4 This is a schematic diagram of the Res-CBAM attention mechanism module structure of the drone image detection method of the embodiment of the present application;

[0047] Figure 5 Schematic diagram of the BiFPN network structure of the drone image detection method according to an embodiment of the present application. DETAILED DESCRIPTION

[0048] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0049] The present application embodiment provides a method for detecting drone images by integrating visible light and infrared. Figure 1 , Figure 2 As shown, including:

[0050] In step S101, visible light sensors and infrared sensors are activated in groups, and drone detection is performed on the target area according to a preset detection strategy. In some embodiments, the visible light sensors and infrared sensors are activated in groups, and drone detection is performed on the target area according to a preset detection strategy, including:

[0051] After initialization, the visible light sensor and infrared sensor are activated, and scanning is performed according to the set radius, scanning adjustment interval and scanning angle. For example, during initialization, the visible light and infrared sensors activated at the same time immediately begin to intensively scan the circular area with a radius of 2 kilometers centered on the operation center. The sensor automatically adjusts the scanning direction every 3 seconds, and each scan covers an angle range of 120 degrees, achieving a 360-degree full-range scan every 9 seconds. In addition, the two cameras simultaneously capture images at a high rate of 15 frames per second. When the visible light camera detects abnormal color changes or movement patterns, the system will immediately compare and fuse the data with the infrared camera. Once a suspicious target is detected, the system switches to enhanced tracking mode. At this time, visible light and infrared images are used for analysis and tracking at the same time to ensure that the target can be effectively monitored in different environments. At the same time, the drone detection system is activated to further analyze the target.

[0052] In step S102, for the detected infrared image and visible light image, feature extraction is performed using independent branches respectively, and a Res-CBAM attention mechanism module is set between the two independent branches. Specifically, in Backbone, two streams are used to independently process infrared and visible light images, and specialized feature extraction is performed respectively. At the same time, the Res-CBAM attention mechanism module is used between the two feature extraction branches to further extract the fusion features of infrared and visible light.

[0053] In step S103, the image features extracted by the two branches are fused using the BiFPN structure. Specifically, the BiFPN structure is used for feature fusion in the neck of the entire network architecture. By using the combined effect of multiple BiFPN layers and the Res-CBAM attention mechanism module, the features are further fused at multiple scales.

[0054] In step S104, based on the fused image features, target detection is performed based on the classification network and the regression network to output the classification result. A decoupled detection head is used in the head part, and the classification and regression tasks are processed independently by two branches to efficiently obtain the final classification result and target frame position, thereby achieving the target detection task.

[0055] This application performs feature extraction of infrared images and visible light images in the Backbone part, using a two-stream backbone, such as Figure 3 As shown, the features of the infrared image and the visible light image are extracted by two independent branches respectively. In some embodiments, for the detected infrared image and the visible light image, the features are extracted by using independent branches respectively, including:

[0056] In the visible light branch for extracting image features of visible light images, the visible light image is input into the stem sampling layer (Stem Layer). The Stem Layer contains three convolutional networks with the same structure, one of which consists of a convolutional layer with a convolution kernel size of 3×3 and a step size of 2, a batch normalization layer and a SiLU function, and the other two convolutional networks consist of a convolutional layer with a convolution kernel size of 3×3 and a step size of 1, a batch normalization layer and a SiLU function. After being processed by the Stem Layer, a feature map with 64 channels is obtained. The feature map processed by the Stem Layer is output through four CSPResLayers with a step size of 2 and the number of channels doubled successively to output a feature map with a set number of channels. In the specific example, the output feature maps after the second, third, and fourth CSPResLayers are F, respectively. rgb1 、F rgb2 and F rgb3 , the number of channels are 256, 512 and 1024 respectively.

[0057] In the infrared image branch that extracts the image features of the infrared image, the infrared image is input into the network with the same structure as that of the visible light image to output a feature map with a set number of channels. The infrared image branch uses the same feature extraction network as the visible light feature extraction branch. The entire branch consists of a Stem Layer and four CSPResLayers. The output feature maps after the second, third, and fourth CSPResLayers are F ir1 、F ir2 and F ir3 , whose channel numbers are 256, 512 and 1024 respectively. After obtaining three sets of visible light feature maps and infrared feature maps with matching channel numbers and sizes, these three sets of feature maps are respectively passed through three Res-CBAM attention mechanism modules to obtain three sets of infrared and visible light fusion features of different sizes.

[0058] F fi =ResCBAM(F rgbi ,F iri )(i=1,2,3)

[0059] Fusion feature F f1 、F f2 and F f3 It will be used as the input of the feature fusion module and applied to subsequent embodiments.

[0060] In some embodiments, Figure 4 As shown, the Res-CBAM attention mechanism module set between two independent branches includes:

[0061] The infrared image features and visible light image features input in groups are concatenated to form the fusion feature F in ;

[0062] The fusion feature F in After a channel attention module is processed to obtain the output features The channel attention module includes a global maximum pooling layer, a global average pooling layer and a shared multi-layer perceptron;

[0063] The fusion feature F in and output features Perform element-wise multiplication to obtain the first output feature

[0064]

[0065] The output features After a spatial attention module, the second output feature is obtained The spatial attention module includes a global maximum pooling layer, a global average pooling layer, two convolutional layers with a convolution kernel size of 3×3 and a Sigmoid function.

[0066] For the first output feature and the second output feature Perform element-wise multiplication to obtain the third output feature

[0067]

[0068] The fusion feature F in and the third output feature Residual connection output to obtain feature map F out ,

[0069]

[0070] In some embodiments, Figure 5 As shown, the image features extracted by the two branches are fused using the BiFPN structure, including:

[0071] In the neck part of the entire network architecture, the fusion feature F of some infrared image features and visible light image features is used. f1 、F f2 and F f3 Further fusion is performed through the BiFPN structure, such as Figure 5 As shown, a Res-CBAM attention mechanism module is provided at the end of the BiFPN structure to process the features.

[0072] The fusion feature F f1 、F f2 and F f3 In the first BiFPBN Block, a top-down network is used for processing, starting from the input fusion feature F f3 After being processed by the first CSPResLayer, the features are obtained

[0073] The characteristics Pass a convolution layer with a convolution kernel size of 1×1 and upsample the features;

[0074] The doubled-size features obtained by upsampling are combined with the fusion features F f2 Perform horizontal splicing operation, and then further process it through the second CSPResLayer to obtain the feature

[0075] The characteristics Through a 1×1 convolution layer and upsampling to the feature size and fusion feature F f1 Same as above, and the processed features are combined with the fusion features F f1 Perform horizontal splicing operation and obtain features through the third CSPResLayer processing

[0076] The characteristics After a convolution layer with a kernel size of 3×3 and a stride of 2, the features with half the size are obtained and combined with the features Perform splicing operations;

[0077] The concatenated features are processed by the fourth CSPResLayer to obtain features The characteristics And the fusion feature F f2 Add and fuse to get features

[0078]

[0079] The characteristics After a convolution layer with a kernel size of 3×3 and a stride of 2, the features with half the size are obtained and combined with Perform the splicing operation and process the spliced ​​features through the fifth CSPResLayer to obtain features The characteristics And the fusion feature F f3 Add and fuse to get features

[0080] The characteristics feature feature As the input of the next BiFPN Block.

[0081] In some embodiments, the image features extracted by the two branches are fused using the BiFPN structure, including:

[0082] In the Neck part, three BiFPN Blocks are used continuously to obtain the output fusion feature. and as well as,

[0083] The output features of the three scales are input into three corresponding Res-CBAM attention mechanism modules respectively to output the output features of the Neck part. and

[0084] In some embodiments, performing target detection based on a classification network and a regression network based on the fused image features includes:

[0085] A decoupled detection head is used in the head part of the network architecture, and the classification network and regression network are input respectively to independently complete the classification task and regression task, thereby achieving the final detection, where:

[0086] In the classification network, the input fusion features are passed through the first convolutional network, which contains two 2D convolutional layers with 3×3 convolution kernels, a 2D batch normalization layer and a SiLU function, and then continue to pass through the second convolutional network, which contains a 2D convolutional layer with 3×3 convolution kernels, a 2D batch normalization layer and a SiLU function. Then, the features are passed through a 2D convolutional layer with a 1×1 convolution kernel, and the final classification result is output. The BCE Loss function is used for training in the classification network.

[0087] In the regression network, the input fusion features are processed using the same structure as the classification network. The main difference in the regression part is that a specific loss function is used for model training. Specifically, the DistributionFocal Loss and CIoU Loss loss functions are used for training in the regression network, which helps to accurately adjust the position and size of the detection box to obtain more accurate regression results. In contrast, the classification task is trained using the BCE Loss function to improve the accuracy of classification. The regression results are obtained through such a regression network.

[0088] The method of the embodiment of the present application designs a two-stream backbone network: by using a two-stream backbone in the backbone to independently process infrared and visible light data, specialized processing and complementarity of data are achieved. This method not only allows each image to be processed by a specialized network architecture according to its characteristics, but also combines the advantages of infrared images in low-light conditions with the advantages of visible light images in providing rich environmental details, thereby greatly improving the accuracy and robustness of drone detection.

[0089] The method of the embodiment of the present application designs the Res-CBAM attention mechanism: the Res-CBAM attention mechanism is introduced to fuse the features from infrared and visible light, and the target feature information in the infrared features is fused with the scene detail information in the visible light features, which enhances the complementarity of the features and improves the overall performance of the model. By focusing on key features and suppressing irrelevant information, the attention mechanism not only enhances the efficiency and interpretability of the detection system, but also improves the adaptability and flexibility of the model to environmental changes and drone features under different conditions.

[0090] An embodiment of the present application also proposes a drone image detection system that integrates visible light and infrared, including a processor and a memory, wherein a computer program is stored on the memory, and when the computer program is executed by the processor, the steps of the drone image detection method that integrates visible light and infrared are implemented as described above.

[0091] It should be noted that in the various embodiments of the present application, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the presence of other identical elements in the process, method, article or device including the element.

[0092] The serial numbers of the above-mentioned embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0093] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for a terminal (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods described in each embodiment of the present application.

[0094] The embodiments of the present application are described above in conjunction with the accompanying drawings, but the present application is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present application, ordinary technicians in this field can also make many forms without departing from the purpose of the present application and the scope of protection of the claims, all of which are within the protection of the present application.

Claims

1. A drone image detection method integrating visible light and infrared, characterized in that: include: Start visible light sensors and infrared sensors in groups and perform drone detection in the target area according to the preset detection strategy; For the detected infrared image and visible light image, independent branches are used for feature extraction respectively, and a Res-CBAM attention mechanism module is set between the two independent branches; The image features extracted by the two branches are fused using the BiFPN structure; Based on the fused image features, target detection is performed based on the classification network and regression network to output the classification results.

2. The method for detecting unmanned aerial vehicle images by integrating visible light and infrared as claimed in claim 1, characterized in that: The visible light sensors and infrared sensors are activated in groups, and drone detection is performed in the target area according to the preset detection strategy, including: After initialization, the visible light sensor and the infrared sensor are activated, and scanning is performed according to the set radius, scanning adjustment interval, and scanning angle.

3. The method for detecting unmanned aerial vehicle images by integrating visible light and infrared as claimed in claim 2, characterized in that: For the detected infrared image and visible light image, feature extraction is performed using independent branches, including: In the visible light branch for extracting image features of the visible light image, the visible light image is input into a stem sampling layer (StemLayer), the Stem Layer comprises three convolutional networks with the same structure, one of which consists of a convolutional layer with a convolutional kernel size of 3×3 and a stride of 2, a batch normalization layer and a SiLU function, and the other two convolutional networks consist of a convolutional layer with a convolutional kernel size of 3×3 and a stride of 1, a batch normalization layer and a SiLU function. The feature map processed by the StemLayer is output through four CSPResLayers with a stride of 2 and a number of channels doubled successively to obtain a feature map with a set number of channels; In the infrared image branch for extracting image features of the infrared image, the infrared image is input into a network with the same structure as that for extracting image features of the visible light image to output a feature map with a set number of channels.

4. The method for detecting unmanned aerial vehicle images by integrating visible light and infrared as claimed in claim 3, characterized in that: The Res-CBAM attention mechanism module set between two independent branches includes: The infrared image features and visible light image features input in groups are concatenated to form the fusion feature F in ; The fusion feature F in After a channel attention module is processed, the output features are obtained The channel attention module includes a global maximum pooling layer, a global average pooling layer and a shared multi-layer perceptron; The fusion feature F in and output features Perform element-wise multiplication to obtain the first output feature The output features After a spatial attention module, the second output feature is obtained The spatial attention module includes a global maximum pooling layer, a global average pooling layer, two convolutional layers with a convolution kernel size of 3×3 and a Sigmoid function; For the first output feature and the second output feature Perform element-wise multiplication to obtain the third output feature The fusion feature F in and the third output feature Residual connection output to obtain feature map F out .

5. The method for detecting unmanned aerial vehicle images by integrating visible light and infrared as claimed in claim 4, characterized in that: The image features extracted by the two branches are fused using the BiFPN structure, including: In the Neck part, the fusion feature F of some infrared image features and visible light image features is f1 、F f2 and F f3 Further fusion is performed through the BiFPN structure, and a Res-CBAM attention mechanism module is set at the end of the BiFPN structure to process the features; The fusion feature F f1 、F f2 and F f3 In the first BiFPBN Block, a top-down network is used for processing, starting from the input fusion feature F f3 After being processed by the first CSPResLayer, the features are obtained The characteristics Pass a convolution layer with a convolution kernel size of 1×1 and upsample the features; The doubled-size features obtained by upsampling are combined with the fusion features F f2 Perform horizontal splicing operation, and then further process it through the second CSPResLayer to obtain the feature The characteristics Through a 1×1 convolution layer and upsampling to the feature size and fusion feature F f1 Same as above, and the processed features are combined with the fusion features F f1 Perform horizontal splicing operation and obtain features through the third CSPResLayer processing The characteristics After a convolution layer with a kernel size of 3×3 and a stride of 2, the features with half the size are obtained and combined with the features Perform splicing operations; The concatenated features are processed by the fourth CSPResLayer to obtain features The characteristics And the fusion feature F f2 Add and fuse to get features The characteristics After a convolution layer with a kernel size of 3×3 and a stride of 2, the features with half the size are obtained and combined with Perform the splicing operation and process the spliced ​​features through the fifth CSPResLayer to obtain features The characteristics And the fusion feature F f3 Add and fuse to get features The characteristics feature feature As the input of the next BiFPN Block.

6. The method for detecting unmanned aerial vehicle images by integrating visible light and infrared as claimed in claim 5, characterized in that: The image features extracted by the two branches are fused using the BiFPN structure, including: In the Neck part, three BiFPN Blocks are used continuously to obtain the output fusion feature. and as well as, The output features of the three scales are input into three corresponding Res-CBAM attention mechanism modules respectively to output the output features of the Neck part. and 7. The method for detecting unmanned aerial vehicle images by integrating visible light and infrared as claimed in claim 5, characterized in that: Based on the fused image features, target detection is performed based on the classification network and regression network, including: In the Head part, a decoupled detection head is used to input the classification network and regression network respectively to independently complete the classification task and regression task, where: In the classification network, the input fusion features are passed through the first convolutional network, which contains two 2D convolutional layers with 3×3 convolution kernels, a 2D batch normalization layer and a SiLU function, and then continue to pass through the second convolutional network, which contains a 2D convolutional layer with 3×3 convolution kernels, a 2D batch normalization layer and a SiLU function. Then, the features are passed through a 2D convolutional layer with a 1×1 convolution kernel, and the final classification results are output. The Distribution Focal Loss and CIoU Loss loss functions are used for training in the classification network. In the regression network, the input fusion features are processed using the same structure as the classification network, and the BCE Loss function is used for training in the regression network to obtain the regression results.

8. A drone image detection system integrating visible light and infrared, characterized in that: It includes a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the steps of the unmanned aerial vehicle image detection method for fusing visible light and infrared are implemented as described in any one of claims 1 to 7.

Images