Target identification method and device

Abstract

The invention discloses a target identification method and device, and the method comprises the steps: S1, obtaining an image data set of air refueling in a complex scene, and enabling the data set to comprise a visible light image and an infrared image; s2, fusing the visible light image and the infrared image to obtain a fused image; s3, training an improved YOLOX model according to the fused image; and S4, inputting a real-time image of air refueling in a complex scene into the trained improved YOLOX model for target recognition. By adopting the technical scheme of the invention, the requirements of keeping high recognition rate and real-time performance during air refueling can be met in complex environments such as daytime, night and the like.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and in particular relates to a target recognition method and device. Background Art

[0002] With the continuous advancement of science and technology, modern warfare has gradually evolved into a contest of military technological strength. As an important part of modern air force operations, aerial refueling has a vital impact on the army's sustained combat capability. In the process of aerial refueling, accurate target identification and precise operational control are key factors to ensure the success of the mission. By fusing infrared and visible light images, high-quality image data of the tanker and the receiver under different time and space conditions can be obtained, which can provide key visual information for pilots or automated systems to help make correct operational decisions in complex flight environments.

[0003] The current technical research on aerial refueling missions faces many challenges, such as the small size of the data set, the difficulty of obtaining samples, and the complexity and variability of flight scenarios. Aerial refueling operations are usually carried out in high-speed, dynamic and high-risk environments. The relative position and attitude of the receiving aircraft and the tanker are constantly changing. There may be interference factors such as clouds and terrain obstructions around. These complex scenes make target recognition more difficult than ordinary flight missions. Existing target recognition algorithms perform well in simple scenarios, but often fail to meet actual needs in complex aerial refueling environments.

[0004] Since the aerial refueling mission has extremely high requirements for real-time performance and accuracy, the existing target recognition methods have the following problems: 1) Although the target recognition algorithm based on the candidate box has high accuracy, the computational redundancy is large, resulting in slow recognition speed; the target recognition algorithm based on regression has fast recognition speed but relatively low accuracy. 2) In complex flight environments, the visible light visibility is low at night, and infrared image information needs to be integrated. 3) The aerial refueling mission requires extremely high real-time performance and accuracy, and many current algorithms find it difficult to strike a balance between the two, resulting in the inability to achieve stable recognition results in highly dynamic environments. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a target recognition method and device that can meet the requirements of maintaining high recognition rate and real-time performance during aerial refueling in complex environments such as day and night.

[0006] To achieve the above object, the present invention adopts the following technical solution:

[0007] A target recognition method, comprising:

[0008] Step S1, obtaining an image dataset of aerial refueling in complex scenes, wherein the dataset includes visible light images and infrared images;

[0009] Step S2, fusing the visible light image and the infrared image to obtain a fused image;

[0010] Step S3, improving the YOLOX model according to the fused image training;

[0011] Step S4: input the real-time image of aerial refueling in complex scenes into the trained improved YOLOX model for target recognition.

[0012] Preferably, in step S1, the image dataset of aerial refueling in complex scenes is annotated and data enhanced.

[0013] Preferably, step S2 comprises:

[0014] The visible light image and the infrared image are filtered and decomposed into two layers by using a mean filter to obtain the base layer and detail layer images respectively;

[0015] Based on the principle of visual saliency, the weight maps of the base layer and the detail layer are obtained, and the RGB and IR images are fused according to the weight information of their respective layers.

[0016] Preferably, in step S3, improving the YOLOX model includes: improving CS-Backbone and improving SilmFPN; wherein, the C2f_SAConv module is added to the improved CS-Backbone, the VoV-GSCSP module is added to the improved SilmFPN, and GFiocaler-iou is used as the loss function for bounding box regression in object detection.

[0017] The present invention also provides a target recognition device, comprising:

[0018] An acquisition module is used to acquire an image dataset of aerial refueling in complex scenes, wherein the dataset includes visible light images and infrared images;

[0019] A fusion module, used for fusing the visible light image and the infrared image to obtain a fused image;

[0020] The training module is used to improve the YOLOX model based on the fused image training;

[0021] The recognition module is used to input real-time images of aerial refueling in complex scenes into the trained improved YOLOX model for target recognition.

[0022] Preferably, the acquisition module is also used to annotate and enhance the image dataset of aerial refueling in complex scenes.

[0023] Preferably, the fusion module includes:

[0024] A first calculation unit is used to filter the visible light image and the infrared image into two layers by using a mean filter to obtain a base layer image and a detail layer image respectively;

[0025] The second calculation unit is used to obtain the weight maps of the base layer and the detail layer based on the principle of visual saliency, and fuse the RGB and IR images according to the weight information of their respective layers.

[0026] Preferably, the improved YOLOX model includes: improved CS-Backbone and improved SilmFPN; wherein, the C2f_SAConv module is added to the improved CS-Backbone, the VoV-GSCSP module is added to the improved SilmFPN, and GFiocaler-iou is used as the loss function for bounding box regression in object detection.

[0027] This paper effectively enhances the capabilities of YOLOX in key links such as feature extraction, feature fusion and loss optimization through image fusion, improved CS-Backbone, improved SilmFPN and GFiocaler-iou loss function design, thereby significantly improving the overall performance of target recognition. This provides important technical support for the application of YOLOX, an efficient target recognition model, in aerial refueling. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] 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 embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0029] Figure 1 is a flow chart of the target identification method of the present invention;

[0030] Figure 2 It is a network structure diagram of the backbone feature extraction of the improved YOLOX algorithm in the present invention;

[0031] Figure 3 It is a diagram of the C2f_SAConv module in CS-Backbone of the present invention;

[0032] Figure 4 It is the Gsconv convolution module diagram in SilmFPN of the present invention;

[0033] Figure 5 It is the VoV-GSCSP feature fusion module in SilmFPN of the present invention. DETAILED DESCRIPTION

[0034] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0035] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0036] Embodiment 1:

[0037] The embodiment of the present invention provides a target recognition method. In order to cope with the complex environment of day and night, the visible light (RGB) and infrared (IR) original images are first fused at the front end to obtain more image information. Image fusion is to filter the RGB and IR images through a mean filter to decompose them into two layers, respectively obtain the base layer and detail layer images, and then obtain the weight maps of the base layer and detail layer based on the principle of visual saliency. The RGB and IR images are fused according to the weight information of their respective layers. After the fused image is obtained, target recognition is performed. The fused image is sent to a trained model, and front-end fusion is selected for target recognition, rather than extracting RGB and IR features separately and then fusing the features for recognition. This is because front-end fusion is easier to implement and has a smaller amount of calculation, and is more in line with the requirements of making decisions quickly and accurately in aerial refueling. In the target recognition link, the fused and reconstructed image will enter the improved CS-Backbone parallel feature extraction backbone network. The backbone extraction network is as follows Figure 2 The network has been added with the C2f_SAConv module. The improved network model improves the adaptability and accuracy of the network in processing features of different scales and complexities, thus showing significant performance improvement in the field of object detection and segmentation. The C2f_SAConv module is shown in the figure. Figure 3 As shown. By building the Neck network of SilmFPN, this module emphasizes the difference between traditional standard convolutions and modern lightweight depth-separable convolutions, where the latter reduces the complexity of computation while maintaining sufficient accuracy, which is especially beneficial for environments with limited computing resources. It is often used in neural network architectures for mobile and edge devices to improve operating efficiency. Finally, the loss function was redesigned to form GFOcaler-iou, which is a loss function used for bounding box regression in object detection. At the same time, it directly minimizes the distance between the center points of the predicted box and the true box to accelerate convergence, which can improve the performance of detectors in different detection tasks and improve the accuracy of target detection.

[0038] Furthermore, an embodiment of the present invention adopts improved YOLOX target recognition. First, a complex scene aerial refueling data set is collected and labeled, and then the data set is expanded using data enhancement, and the data set is divided into a training set and a validation set; the training set is used to train the YOLOX model to obtain a trained improved YOLOX model for complex scene aerial refueling target recognition. The specific method is: the aerial refueling aircraft and drogue data set is obtained through the camera radar, the unmanned vehicle and the constructed airsim simulation platform, the constructed basic data set is preliminarily labeled using the pre-trained model, and then manually refined and corrected. Then, data enhancement methods are used, including translation, cropping, rotation, mirroring, and scale transformation, to expand the insufficient number of category image data in the labeled data set; the sample data set is divided into a training set and a validation set, and the fused image is input into the CS-Backbone parallel feature extraction backbone network module. The fused image is extracted through the backbone network to obtain feature maps of three different scales with output sizes of 80×80×256, 40×40×512, and 20×20×1024. The entire network gradually reduces the spatial size of the feature map by downsampling, and increases the number of channels to capture richer features. During the entire aerial refueling process, the position change of the distant receiving aircraft is not obvious, but at close range, the scale of the receiving aircraft in the camera field of view changes greatly, which interferes with recognition. The selection of the CS-Backbone network structure can effectively process complex feature patterns through this "double observation and combination" strategy, especially in the case of large scale changes. Then, the three different feature maps are input into SilmFPN for feature fusion. The VoV-GSCSP module in SilmFPN is used to speed up the prediction calculation of the image, and the feature pyramid network (FPN) is combined with the path aggregation network (PAN) to enhance the expression ability of multi-scale features. In this way, the high accuracy is maintained during the operation of the aerial refueling aircraft while reducing the computational complexity and reasoning time, reducing the hardware requirements of the aircraft, and meeting the high real-time requirements in the aerial refueling environment. The fused feature maps of three different scales are obtained, which will be input into the head network of YOLO for the final target detection task. For each scale of feature maps, the head network will generate the original prediction value through convolution. These prediction values ​​contain the location information (bounding box coordinates), confidence score (ConfidenceScore) and category probability of each preset anchor box (anchorbox). The original prediction value needs to be decoded into the actual bounding box coordinates and category probability. For bounding box coordinates, this usually involves converting the predicted value into an offset relative to the grid cell and then into absolute coordinates. After obtaining the bounding boxes and confidence scores of all predicted boxes, YOLO usually uses a non-maximum suppression algorithm to filter out overlapping detection results, retaining only those bounding boxes with the highest confidence as the final detection results.In training mode, the GFiocaler-iou loss function is used to calculate the regression loss, classification loss, and object loss based on the true label and the model's prediction, and weight them together into a total loss to optimize the model parameters so that the model can better learn the mapping relationship from the input image to the target detection result. During the training process, after each forward propagation, the model weights are updated according to the calculated loss, thereby gradually improving the performance of the model. This design allows the loss function to be sensitive to the IOU value within a certain range, which helps the model better learn to extract features from moderately difficult samples, rather than just focusing on the easiest or most difficult samples, so as to train the model to meet the requirements of accurate and high-precision recognition of the receiving aircraft and the drogue during aerial refueling.

[0039] CS-Backbone uses switchable atrous convolution (SAC), which is an advanced convolution mechanism for enhancing feature extraction in object detection and segmentation tasks. The core idea is to apply different atrous rates to the same input features for convolution. Atrous convolution expands the receptive field by introducing additional space (i.e., holes) in the convolution kernel without increasing the number of parameters or the amount of computation, and can also capture features of different scales. Another key feature is the use of switching functions to combine the results of convolutions with different atrous rates. These switching functions are spatially dependent, meaning that each position of the feature map may have a different switch to control the output, making the network more flexible to the size and scale of the features. The conversion of a traditional convolutional layer to a SAC layer is achieved by using the same weights (except for a trainable difference) in convolution operations with different atrous rates. This conversion mechanism includes an average pooling layer and a 1×1 convolution layer to implement the switch function, a design that reduces model complexity while maintaining flexibility. SAC has two global context modules before and after, which add image-level information to the features to help the network better understand and process the overall content of the image, thereby improving the quality and accuracy of feature extraction and enabling it to work effectively in a wider context. The backbone network extracts features from the fused image, and in this process, feature information of different scales will be generated through convolution operations. The structure of SAC replaces the convolution operation, allowing it to observe the features of different scales in the backbone network twice, but using different void rates each time. This means that the same scale feature will be processed by two different configurations of convolution kernels, each of which corresponds to a specific void rate. Doing so can capture more feature information of different scales, thereby more comprehensively understanding and analyzing the input data. The output results obtained with different void rates are then combined together through switch functions. These switches determine how to select or fuse information from the two convolutions to generate the final output features. The operation of the switch may depend on the characteristics of the feature itself, such as its spatial position. Through this "double observation and combination" strategy, SAC can effectively handle complex feature patterns, especially when the scale changes are large. This approach not only improves the flexibility and adaptability of feature extraction, but also improves the accuracy and efficiency in object detection and segmentation tasks.

[0040] SilmFPN is designed to speed up prediction. The input image almost always undergoes a similar transformation process on the backbone, which is to gradually transfer spatial information to the channel. Each compression of the spatial dimension (width and height) of the feature and the expansion of the channel dimension will cause the loss of some semantic information. The channel dense convolution maximizes the retention of hidden connections between each channel, while the channel sparse convolution completely cuts off these connections. The GSConv module retains these connections as much as possible with lower time complexity. The GSConv module is shown in the figure. Figure 4 As shown in Figure 2, after GSConv, the GS bottleneck (GSbottleneck) and cross-stage partial network (GSCSP) modules VoV-GSCSP are introduced. The VoV-GSCSP module is as follows: Figure 5 As shown in the figure, these modules are designed to further improve performance. In practical applications, simpler structural modules are more likely to be adopted because they are easier to implement in hardware. GSConv uses shuffle to infiltrate the features generated by SC (channel dense convolution) into each part of the features generated by DSC. Shuffle is a uniform mixture that allows information from SC to be fully mixed into the output of DSC by evenly exchanging local feature information on different channels. The GSbottleneck module is to further enhance the network's ability to process features and improve the model's learning ability by stacking GSConv modules. The VOV-GSCSP module is an efficient cross-stage partial network (CSP) module designed using a one-time aggregation strategy, using different structural design schemes to improve feature utilization efficiency and network performance. These modules are designed to reduce computational complexity and inference time while maintaining accuracy. Through such a modular design, a network architecture suitable for specific tasks can be flexibly constructed as needed.

[0041] GFocaler-iou is a loss function for bounding box regression in object detection. It improves the performance of detectors in different detection tasks by focusing on different regression samples. This is done by reconstructing the IOU loss through linear interval mapping to achieve attention to different samples.

[0042] The impact of the distribution of difficult and simple samples in bounding box regression on the regression results is analyzed and considered, which is an aspect that is often overlooked in the traditional IOU loss function. GFocaler-iou makes up for the shortcomings of existing bounding box regression methods through its unique method, and can directly minimize the distance between the center points of the predicted box and the true box to accelerate convergence, thereby further improving the detection performance in different detection tasks. Each feature map finally converts the channel into detection information output through 1×1 convolution, completes the prediction of target category and confidence, and decodes and suppresses the output of detection information. On this basis, the final detection result is obtained, and then the total loss is calculated and the gradient is back-transferred through the improved loss function to complete the network parameter update and optimize the final target detection algorithm performance. When the IOU is less than a lower threshold d, the loss is 0; when the IOU is greater than an upper threshold u, the loss is 1; and when the IOU is between d and u, the loss is a function that increases linearly according to the IOU value. This design allows the loss function to be sensitive to the IOU value within a certain range, so that it can focus more on samples with a moderate overlap between the predicted bounding box and the true bounding box, that is, samples that are neither too difficult nor too easy. This helps the model better learn to extract features from samples of moderate difficulty, rather than just focusing on the easiest or most difficult samples. At the same time, an additional metric is introduced to solve this problem, namely the concept of the minimum enclosing box. This not only considers the intersection and union, but also the spatial relationship between the two boxes.

[0043] like Figure 1 As shown, an embodiment of the present invention provides a target recognition method, including:

[0044] Step S1, obtaining an image dataset of aerial refueling in complex scenes, wherein the dataset includes visible light images and infrared images;

[0045] Step S2, fusing the visible light image and the infrared image to obtain a fused image;

[0046] Step S3, improving the YOLOX model according to the fused image training;

[0047] Step S4: input the real-time image of aerial refueling in complex scenes into the trained improved YOLOX model for target recognition.

[0048] As an implementation method of an embodiment of the present invention, in step S1, the collection and annotation of the aerial refueling data set in a complex scene can optionally autonomously obtain the aircraft and drogue data set for aerial refueling through camera radar, drone and the constructed airsim simulation platform; then adjust each image to have the same size information to obtain several pre-processed target images; finally, use the LabelImg annotation tool to annotate the pre-processed images.

[0049] As an implementation method of an embodiment of the present invention, in step S2, for two registered source images of the same size (visible light and infrared images), a mean filter is first used to decompose them into a base layer containing large-scale changes and a detail layer containing small-scale changes. Next, it is necessary to detect the visually salient areas in the source image and obtain the image saliency map according to the formula ξ(x, y) = ||φ μ (x,y)-φ η (x,y)||where|| is the Euclidean distance. φ μ is the output of the averaging filter. η is the output of the median filter. According to the results of visual saliency detection, a weight map is constructed. The specific process is as follows: The role of the weight map is to assign different weights to different detail layers during the fusion process to ensure that important information is retained first. Different fusion strategies are adopted for different layers. For the base layer, the average fusion rule is used to fuse the base layer. The specific process is as follows: For the detail layer, fusion is performed through weights. The specific process is as follows: For the fused detail layer and base layer, the two are fused to obtain the final reconstructed fused image. The specific process is as follows: γ(x,y)=φ B (x,y)+φ D (x,y).

[0050] As an implementation method of an embodiment of the present invention, in step S3, data enhancement is used to expand the military target data set, and the data set is divided into a training set and a verification set; optionally, since there is a problem that the number of images of a certain category is small in the data set, the present invention first uses traditional data enhancement methods, including translation, cropping, rotation, mirroring, and scale transformation to simply expand the image data of a certain category with a small number in the sample data set to increase the diversity of training samples; finally, the present invention divides 8012 images into a training set and a verification set in a ratio of 7:3.

[0051] Furthermore, the training set is used to train the improved YOLOX model to obtain the trained improved YOLOX model; the fused image is input into the CS-Backbone parallel feature extraction backbone network module, and the fused image is extracted through the backbone network. The image first passes through the Focus layer. Focus is a special convolution pre-processing method that improves the information representation capability by dividing the input spatial dimensions (W, H) into blocks and converting them into channel dimensions. For a 640×640×3 input, the Focus layer converts it into a feature map with a shape of 320×320×64. The number of channels here becomes 64, because the Focus layer extracts information from different positions in the spatial dimension and combines it into more channels, and then enters the dark2 module, first convolves through BaseConv to increase the number of channels from 64 to 128, and downsamples through convolution with a step size of 2, and the feature map size is reduced to 160×160×128. Then enter C2f_SAConv, replace the original CrossStagePartial (CSP) structure with C2f and introduce the SAConv module. This method enables the network to adapt to features of different scales more flexibly, so as to more accurately identify and segment objects in the image. This layer will further process the feature map, but will not change the size. It only operates in the channel dimension, and the output is still 160×160×128. Then enter the dark3 and dark4 modules in turn, and convolve through BaseConv to increase the number of channels from 128 to 256 (the dark4 module increases to 512), and downsample through convolution with a step size of 2, and the feature map size is reduced to 80×80×256 (the dark4 module is 40×40×512), and then enter CSPLayer, which contains a residual module of n=base_depth*3 (the dark4 module is the same here), which allows the network to extract features at a deeper level. The output feature map of dark3 is 80×80×256 (the output feature map of dark4 is 40×40×512). Finally, the dark5 module is convolved through BaseConv to increase the number of channels from 512 to 1024, and the feature map size is reduced to 20×20×1024 through convolution downsampling with a step size of 2. Then it enters the SPPBottleneck layer (SPP pyramid pooling) to further enhance the global receptive field of the feature map. Finally, a CSPLayer is used to further extract features and output the final feature maps of 80×80×256, 40×40×512, and 20×20×1024. The backbone network extracts multiple feature maps with different semantics and scales from the input image through a series of convolution and pooling operations.The Neck network adopts a bidirectional information transmission strategy, combining top-down and bottom-up feature paths to effectively fuse feature information of different scales and generate a fused feature map. The specific operations are as follows: F1, F2 and F3 correspond to feature maps of 20×20×1024, 40×40×512, and 80×80×256 scales, respectively. First, the channel of F1 is compressed by the lateral_conv0 operation, and the 1024 channels are compressed to 512 channels, and the feature map is still 20×20×512. Then the feature map is upsampled by the upsample operation, and the size becomes 40×40×512 (upsampled to the size of F2). The upsampled feature map is concat with F2 to obtain a feature map of 40×40×1024. The concatenated feature map passes through the VovgscspP4 layer and outputs a feature map of 40×40×512. In this way, by integrating the features from the deep layer F1, high-level semantic information is supplemented. The feature map output by VovgscspP4 is channel compressed through reduce_conv1, and 512 channels are compressed to 256 channels, and the output is still 40×40×256. Then the feature map is upsampled to 80×80×256 (upsampled to F3 size). Then, the upsampled feature map is concat with F3 to obtain a feature map of 80×80×512. The concatenated feature maps are fused through the VovgscspP3 layer, and a feature map of 80×80×256 is output. At this time, the first fused feature map pan_out2 is generated. Pan_out2 is convoluted with bu_conv2 for 3×3, and the size is downsampled to 40×40×256, which is concatenated with fpn_out1 (40×40×256) to obtain a feature map of 40×40×512. Then it is processed through the VovgscspN3 layer, and a feature map of 40×40×512 is output, obtaining pan_out1. pan_out1 is downsampled again through bu_conv1 to a size of 20×20×512, and concatenated with fpn_out0 (20×20×512) output by FPN to obtain a 20×20×1024 feature map. Finally, it is processed through the VovgscspN4 layer to output a 20×20×1024 feature map to obtain pan_out0 in the network. pan_out0, pan_out1, and pan_out2 are sent to the detection head environment as the final output of Neck for the final detection task.The specific operation is as follows: For each input feature map, it first passes through a BaseConv layer (i.e. one of the stems) to unify its channel number to int(256*width). For each preprocessed feature map, the classification branch (cls_conv) will pass through two consecutive convolutional layers (each is a 3x3 convolution with an act activation function) to further extract features for classification. The classification prediction layer (cls_preds) is a 1x1 convolutional layer that maps the features of each position to the category prediction. The output shape is [self.n_anchors*self.num_classes]. Similarly, each feature map will also enter a regression branch (reg_conv) and extract features for bounding box regression through two consecutive 3x3 convolutional layers. The regression prediction layer (reg_preds) is a 1x1 convolutional layer used to predict the bounding box coordinates at each position. The output shape is [4], representing (dx, dy, dw, dh) respectively. The objectness branch (obj_preds) is also a 1x1 convolutional layer used to predict the objectness score of each position. The shape of the output is [self.n_anchors*1]. If in training mode, concatenate reg_output, obj_output, and cls_output according to the channel dimension to form a complete prediction tensor. If in inference mode, concatenate reg_output, obj_output.sigmoid(), and cls_output.sigmoid() according to the channel dimension, where the .sigmoid() function converts the score to a probability. If in training mode, the get_losses method is also called to calculate the loss. This step calculates the regression loss, classification loss, and objectness loss based on the true label and the prediction result. If in inference mode, the output is flattened and rearranged to form a tensor of shape [batch,n_anchors_all,85], where 85 includes 4 regression parameters, 1 objectness score, and num_classes class probabilities. If decode_in_inference is True, the decode_outputs method is also called to decode the output bounding box coordinates from relative form to absolute form. The detection head adopts the bounding box regression technology based on GFOcaler-iou, combined with the loss function, to accelerate the convergence speed of the model and alleviate the limitation of the IOU loss term, thereby improving the accuracy of target detection.After filtering out most of the detection boxes based on the confidence, there are often multiple redundant detection results left around the target to be detected. At this time, the non-maximum suppression (NMS) algorithm is needed to further prune them. Considering the difficulty of regression of small targets in the regression branch, regression loss is introduced to accelerate the convergence speed of the model and reduce the limitation of the IOU loss term. The specific loss function of GFocaler-iou is as follows:.

[0052]

[0053] By calculating the total loss and returning the gradient, the network parameters of the two modules are updated based on the gradient descent method to optimize the final target detection algorithm performance.

[0054] Compared with the prior art, the present invention has the following effective benefits:

[0055] 1. In view of the fact that aerial refueling is carried out in high-speed, dynamic and complex environments, the layered fusion and visual saliency weighting methods are used to fuse and reconstruct visible light and infrared images, realize the information complementarity of the two types of image data, improve the target detection accuracy, and help overcome the problem of missing image information and occlusion in complex environments, thereby interfering with target recognition.

[0056] 2. In order to solve the problem that the feature extraction capability of the YOLOX backbone network is not flexible enough, the CS-Backbone module was redesigned to improve the adaptability and accuracy of the network in processing features of different scales and complexities, thereby solving the large scale changes in close distances during aerial refueling and showing significant performance improvements in the field of target recognition.

[0057] 3. In the feature fusion part, the SilmFPN feature fusion module is introduced, and Gsconv achieves an excellent balance between the accuracy and speed of the model, reducing the burden of the model while maintaining accuracy, helping to reduce the power consumption of computers in aerial refueling aircraft and reduce hardware requirements.

[0058] 4. The loss function has been redesigned. The design of GFiocaler-iou combines the advantages of Focaler-IOU and GIOU, giving the loss function the ability to control the IOU value to a certain extent, so that it can focus more on processing samples with a medium degree of overlap between the bounding box and the true value, that is, samples that are neither too difficult nor too easy. This helps the model better learn to extract features from samples of medium difficulty, rather than just focusing on the easiest or most difficult samples.

[0059] Embodiment 2:

[0060] The embodiment of the present invention further provides a target recognition device, comprising:

[0061] An acquisition module is used to acquire an image dataset of aerial refueling in complex scenes, wherein the dataset includes visible light images and infrared images;

[0062] A fusion module, used for fusing the visible light image and the infrared image to obtain a fused image;

[0063] The training module is used to improve the YOLOX model based on the fused image training;

[0064] The recognition module is used to input real-time images of aerial refueling in complex scenes into the trained improved YOLOX model for target recognition.

[0065] Preferably, the acquisition module is also used to annotate and enhance the image dataset of aerial refueling in complex scenes.

[0066] As an implementation of the embodiment of the present invention, the fusion module includes:

[0067] A first calculation unit is used to filter the visible light image and the infrared image into two layers by using a mean filter to obtain a base layer image and a detail layer image respectively;

[0068] The second calculation unit is used to obtain the weight maps of the base layer and the detail layer based on the principle of visual saliency, and fuse the RGB and IR images according to the weight information of their respective layers.

[0069] As an implementation method of an embodiment of the present invention, the improved YOLOX model includes: an improved CS-Backbone and an improved SilmFPN; wherein, a C2f_SAConv module is added to the improved CS-Backbone, a VoV-GSCSP module is added to the improved SilmFPN, and GFiocaler-iou is used as the loss function for bounding box regression in object detection.

[0070] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A target recognition method, characterized in that: include: Step S1, obtaining an image dataset of aerial refueling in complex scenes, wherein the dataset includes visible light images and infrared images; Step S2, fusing the visible light image and the infrared image to obtain a fused image; Step S3, improving the YOLOX model according to the fused image training; Step S4: input the real-time image of aerial refueling in complex scenes into the trained improved YOLOX model for target recognition.

2. The target recognition method according to claim 1, characterized in that: In step S1, the image dataset of aerial refueling in complex scenes is labeled and enhanced.

3. The target recognition method according to claim 2, characterized in that: Step S2 includes: The visible light image and the infrared image are filtered and decomposed into two layers by using a mean filter to obtain the base layer and detail layer images respectively; Based on the principle of visual saliency, the weight maps of the base layer and the detail layer are obtained, and the RGB and IR images are fused according to the weight information of their respective layers.

4. The target recognition method according to claim 3, characterized in that: In step S3, improving the YOLOX model includes: improving CS-Backbone and improving SilmFPN; wherein, the C2f_SAConv module is added to the improved CS-Backbone, the VoV-GSCSP module is added to the improved SilmFPN, and GFiocaler-iou is used as the loss function for bounding box regression in object detection.

5. A target recognition device, characterized in that: include: An acquisition module is used to acquire an image dataset of aerial refueling in complex scenes, wherein the dataset includes visible light images and infrared images; A fusion module, used for fusing the visible light image and the infrared image to obtain a fused image; The training module is used to improve the YOLOX model based on the fused image training; The recognition module is used to input real-time images of aerial refueling in complex scenes into the trained improved YOLOX model for target recognition.

6. The target recognition device according to claim 5, characterized in that: The acquisition module is also used to annotate and enhance the image dataset of aerial refueling in complex scenes.

7. The target recognition device according to claim 6, characterized in that: The fusion module includes: A first calculation unit is used to filter the visible light image and the infrared image into two layers by using a mean filter to obtain a base layer image and a detail layer image respectively; The second calculation unit is used to obtain the weight maps of the base layer and the detail layer based on the principle of visual saliency, and fuse the RGB and IR images according to the weight information of their respective layers.

8. The target recognition device according to claim 7, characterized in that: The improved YOLOX model includes: improved CS-Backbone and improved SilmFPN; among them, the C2f_SAConv module is added to the improved CS-Backbone, the VoV-GSCSP module is added to the improved SilmFPN, and GFiocaler-iou is used as the loss function for bounding box regression in object detection.

Images