Unmanned aerial vehicle target detection method based on YOLO-LiteMax neural network

Abstract

The invention discloses an unmanned aerial vehicle target detection method based on a YOLO-LiteMax neural network, and the method carries out the recognition of dense small targets under the condition of not increasing model parameters, and comprises the steps: replacing Bottleneck in a C2f module with a FasterNet block through selecting a convolution module in a backbone network, and reducing the redundant information of a feature map; performing differentiation processing on the feature maps of different scales in the neck network, and then highlighting small target information by increasing a large-scale feature map to obtain local details of a small target; and further extracting features in the head network by using group normalization and shared convolution, and identifying dense small targets in the target object picture. According to the unmanned aerial vehicle target detection method, the detection capability of the unmanned aerial vehicle on dense small target objects is effectively enhanced, the detection precision is improved, and the robustness of the unmanned aerial vehicle in a complex environment is enhanced while the efficiency is kept.

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning target detection, and in particular to a drone target detection method based on a YOLO-LiteMax neural network. Background Art

[0002] Object detection algorithms can be roughly divided into two-stage detectors and single-stage detectors. Two-stage detectors, such as the R-CNN (Region-CNN) series, first generate candidate regions and then perform detailed classification and regression. These methods are highly accurate and particularly suitable for complex scenes, but suffer from slow inference speed and high computational effort. In contrast, single-stage detectors like YOLO and SSD combine region proposals and classification into a single step, directly predicting object categories and bounding boxes in images. They are fast, simple in structure, and well suited for real-time detection scenarios, although they often have difficulty detecting small objects and dealing with complex backgrounds.

[0003] The YOLOv8 neural network inherits the core idea of ​​the YOLO series, that is, end-to-end target detection, and simultaneously locates (bounding box regression) and classifies (category prediction) the target through a single regional convolutional neural network. The target detection process of YOLOv8 includes the following key steps: preprocessing the input image, using the YOLOv8 backbone network to extract features in the image, generating the target detection area through convolution operations and anchor boxes, predicting the category of the target in each candidate box and predicting the position offset of each candidate box (i.e. bounding box regression) to determine the position and size of the target, using the NMS algorithm to remove redundant boxes, and outputting the results.

[0004] Compared with the traditional YOLOv7, the YOLOv8 neural network has significantly improved in accuracy. YOLOv8 has shown high accuracy on common target detection datasets such as COCO and VOC. However, in terms of drone object detection, the existing YOLOv8 does not work well on small target detection objects, and there will be false detections and missed detections. Further research is needed to improve the accuracy of small target detection, especially the detection capabilities in scenarios where multiple small targets overlap and are diverse. Summary of the invention

[0005] The problem to be solved by the present invention is to provide a UAV target detection method based on the YOLO-LiteMax neural network, which can enhance the detection capability of dense small target objects and improve the detection accuracy without increasing the number of model parameters, and enhance its robustness in complex environments while maintaining efficiency.

[0006] The present invention adopts the following technical solution: a method for detecting drone targets based on a YOLO-LiteMax neural network, comprising the following steps: Step 1: Build a YOLO-LiteMax target detection model based on YOLOv8 and input training set images for model training; Step 2: The input target object image is subjected to feature extraction through the backbone network to generate feature maps of different scales; the backbone network includes several CBS modules, several selection convolution blocks, and an SPPF module; the selection convolution module replaces the Bottleneck in the C2f module with the FasterNet block to reduce redundant information in the feature map; Step 3: Input the generated feature maps of different scales into the neck network, and use the scale sequence to perform feature fusion; the neck network performs differential processing on the feature maps of different scales, fuses the local details and global information of the multi-scale feature maps, and then highlights the small target information by adding large-scale feature maps, and uses the feature extraction module to obtain the local details of the small target; Step 4: Input the feature map after the neck network features are fused into the head network for classification and regression tasks; in the head network, the detection head uses group normalization and shared convolution to further extract features, share information between features of different scales, and train the classification and regression tasks separately through the decoupling head, and use the scale factor to scale the feature map, and finally obtain the YOLO-LiteMax target detection model that accurately identifies dense small targets; Step 5: Input the drone image to be detected into the trained YOLO-LiteMax target detection model to perform target detection and identify dense small targets in the image.

[0007] Preferably, in step 2, each of the CBS modules is composed of a convolution kernel of size 3 and stride 2, a batch normalization layer and a SiLU activation function; Each of the selected convolution blocks includes a CBS module, a Split layer, and three sequentially connected FasterNet blocks; each FasterNet block performs convolution on only part of the channels, and the output of each FasterNet block is used as the input of the next FasterNet block and stored in a list. After three FasterNet blocks, all feature maps in the list are spliced ​​by channel; The SPPF module consists of CBS and Maxpool operations, which enhances the feature extraction capability of the backbone network by pooling information at multiple scales.

[0008] Preferably, in step 2, in the backbone network, the target object image is subjected to feature extraction through the backbone network, and the method is as follows: Step 201: The input image is downsampled twice by two CBS modules in the zeroth layer and the first layer to obtain a 4-fold downsampled feature map, and then enters the second layer to select the convolution block for convolution; In the second layer of convolutional blocks, the 4x downsampled feature map is first processed by a CBS module and the output enters a Split layer. The Split layer divides the feature map into two feature maps with the same number of channels, and then passes through three FasterNet blocks for feature extraction to obtain a large-size feature map P2 after 4x downsampling. The feature map P2 after 4 times downsampling is passed through the third layer CBS module, downsampled 8 times, and then input into the fourth layer selection convolution block for feature extraction to obtain the feature map P3 after 8 times downsampling; After passing through the fifth layer CBS module, it is downsampled 16 times and then input into the sixth layer selection convolution block for feature extraction to obtain the feature map P4 after 16 times downsampling; After passing through the seventh layer CBS module, it is downsampled 32 times and then input into the eighth layer selection convolution block for feature extraction to obtain the feature map P5 after 32 times downsampling; the feature map P5 is passed through the ninth layer SPPF module to improve the ability of the YOLO-LiteMax target detection network to capture context details; Step 202: Output the 8-fold downsampled feature map P3, the 16-fold downsampled feature map P4, the 32-fold downsampled feature map P5, and the 4-fold downsampled feature map P2 after feature extraction to the Neck network.

[0009] Preferably, in step 3, in the neck network, feature fusion is performed on feature maps of different scales in the following manner: Step 301, modify the connection mode between the backbone network and the neck network, connect the fourth, sixth, and ninth layers of the traditional YOLOv8 backbone network to the neck network, and optimize the backbone network to connect the second, fourth, sixth, and ninth layers to the neck network; wherein, the feature map P5 output by the ninth-layer SPPF module is first convolved by the tenth-layer CBS module after entering the neck network; Step 302: The feature maps P3, P4 and the feature map P5 after convolution processing are used as large, medium and small size feature maps, and are input into the first SAC module for feature fusion, and feature maps of different scales are differentiated to balance the detail retention and global information extraction of the feature maps, so as to avoid the loss of small target features during upsampling and enhance the ability to capture the detail information of small targets; The feature map output by the first SAC module is processed again by a selective convolution block, and the feature map output by the selective convolution block and the feature maps P2 and P3 are input to the second SAC module again to further fully fuse the features, and then processed by the selective convolution block to obtain the P3 scale feature map for use by the P3 detection head; The feature map output by the second SAC module is downsampled through a CBS module consisting of a convolution kernel of size 3 and a stride of 2. The downsampled result is concat-operated with the feature map fused by the first SAC module. After processing by selecting a convolution block, a P4 scale feature map is obtained for use by the P4 detection head. The obtained feature map is downsampled again through a CBS module consisting of a convolution kernel of size 3 and a stride of 2. The downsampled result is concat-operated with the feature map output by the tenth layer of the CBS module. After processing by the selected convolution block, the P5 scale feature map is obtained for use by the P5 detection head. Step 303: Add two C3DSF feature extraction modules to the neck network to capture the correlation between feature maps across scales, input feature maps P3, P4, and P5 into the first C3DSF feature extraction module to fully extract feature map information; Step 304, add the features extracted by the first C3DSF feature extraction module and the feature map fused by the second SAC module, add an upsampling module and a Concat module, fuse the P3 scale feature map by upsampling, and perform a Concat operation with the output of the second layer selected convolution block in the backbone network. After being processed again by the selected convolution block, the feature maps P2, P3, and P4 are input into the second C3DSF feature extraction module for feature extraction, and then add the features processed by the selected convolution block to obtain the P2 scale feature map for use by the P2 detection head.

[0010] Preferably, in the SAC module, adaptive maximum pooling and adaptive average pooling are performed on the large-size feature map, the results of the two poolings are added, and an interpolation upsampling operation is performed on the small-size feature map, and then the processed large, medium, and small-size feature maps are spliced ​​in the channel dimension.

[0011] Preferably, in the C3DSF feature extraction module, the large, medium, and small-sized feature maps are adjusted to the same number of channels, and the medium and small-sized feature maps are subjected to interpolation upsampling operations to adjust the feature map size to the size of the large-sized feature map. The dimensions of the large, medium, and small-sized feature map tensors are expanded and spliced, and 3D convolution is used to extract features from the spliced ​​feature maps. Then, after passing through the normalization layer, the activation function layer, and the maximum pooling layer, the processing results are compressed to the dimensions of the input.

[0012] Preferably, in step 4, in the head network, classification and regression tasks are performed as follows: Step 401, modify the connection mode between the neck network and the head network, add a P2 scale detection head, connect the three-layer scale feature map of the traditional YOLOv8 neck network P3, P4, and P5 to the head network, and modify it to connect the four-layer scale feature map of the neck network P2, P3, P4, and P5 to the head network; Step 402: Integrate group normalization into the head network CBS module, replace BatchNormalization with GroupNormalization, and obtain an improved CGS module; Step 403, the P2 scale feature map is first subjected to further feature extraction by an independent CGS module, and then processed by two CGS modules whose parameters are shared by the P2, P3, P4, and P5 scale feature maps. The features output by the CGS module are respectively output to the classification head and regression head of the 1×1 convolution shared by P2, P3, P4, and P5 for decoupled learning, and the regression features are scaled by the Scale factor, and the Bbox_loss and Cls_loss of the P2 scale feature map are respectively calculated; Step 404, the P3 scale feature map is first subjected to further feature extraction by an independent CGS module, and then processed by two CGS modules whose parameters are shared by the P2, P3, P4, and P5 scale feature maps. The features output by the CGS module are respectively output to the classification head and regression head of the 1×1 convolution shared by P2, P3, P4, and P5 for decoupled learning, and the regression features are scaled by the Scale factor, and the Bbox_loss and Cls_loss of the P3 scale feature map are respectively calculated; Step 405, the P4 scale feature map is first subjected to further feature extraction by an independent CGS module, and then processed by two CGS modules whose parameters are shared by the P2, P3, P4, and P5 scale feature maps. The features output by the CGS module are respectively output to the classification head and regression head of the 1×1 convolution shared by P2, P3, P4, and P5 for decoupled learning, and the regression features are scaled by the Scale factor, and the Bbox_loss and Cls_loss of the P4 scale feature map are respectively calculated; Step 406, the P5 scale feature map is first further extracted by an independent CGS module, and then processed by two CGS modules whose parameters are shared by the P2, P3, P4, and P5 scale feature maps. The features output by the CGS module are respectively output to the classification head and regression head of the 1×1 convolution shared by P2, P3, P4, and P5 for decoupled learning, and the regression features are scaled by the Scale factor, and the Bbox_loss and Cls_loss of the P5 scale feature map are calculated respectively.

[0013] Preferably, the Bbox_loss and Cls_loss of the obtained P2, P3, P4, and P5 scale feature maps are used. The regression loss Bbox_loss is used to optimize the position and size of the predicted bounding box to ensure that the object is accurately located; the classification loss Cls_loss is used to improve the accuracy of category recognition to ensure that the object category is correctly classified. By minimizing these two losses at the same time, the YOLO-LiteMax target detection network achieves a balance between positioning and classification. By building and applying the CGS module in the head network, the feature consistency and detection performance of the YOLO-LiteMax target detection network in multi-scale target detection are enhanced. Through collaborative learning, the detection heads of different scales are jointly optimized during training to improve the robustness and generalization ability of feature extraction.

[0014] The technical solution of the present invention also provides: an electronic device, comprising: one or more processors; a storage device having one or more programs stored thereon; When the one or more programs are executed by the one or more processors, the one or more processors implement any of the above-mentioned drone target detection methods based on the YOLO-LiteMax neural network.

[0015] The technical solution of the present invention also provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps in any of the above-mentioned drone target detection methods based on the YOLO-LiteMax neural network are implemented.

[0016] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects: 1. The UAV target detection method based on the YOLO-LiteMax neural network of the present invention replaces the Bottleneck in the C2f module with the FasterNet block by selecting the convolution module in the backbone network without increasing the model parameters, thereby reducing the redundant information of the feature map; the feature maps of different scales are differentiated in the neck network, and the small target information is highlighted by adding a large-scale feature map to obtain the local details of the small target; the group normalization and shared convolution are used in the head network to further extract features and identify dense small targets in the target object picture.

[0017] 2. The UAV target detection method based on the YOLO-LiteMax neural network of the present invention effectively enhances the detection capability of dense small target objects, improves the detection accuracy, and enhances its robustness in complex environments while maintaining efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is the overall structure diagram of the drone target detection method based on the YOLO-LiteMax neural network of the present invention; Figure 2 This is a structural block diagram of the SCB module of the present invention; Figure 3 The structure diagram of the SAC and C3DSF modules of the present invention; Figure 4 This is a structural diagram of the improved detection head of the present invention. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the application is further elaborated in detail below in conjunction with the accompanying drawings. The described embodiments are only a part of the embodiments involved in the present invention. All non-innovative embodiments of other researchers in the field on this embodiment belong to the protection scope of the present invention. At the same time, for the step numbering in the embodiment of the present invention, it is only set for the convenience of explanation, and the order between the steps is not limited in any way. The execution order of each step in the embodiment can be adaptively adjusted according to the understanding of those skilled in the art.

[0020] In one embodiment of the present invention, a drone target detection method based on a YOLO-LiteMax neural network is provided. Figure 1 As shown, the following steps are included: Step 1: Build a YOLO-LiteMax target detection network based on YOLOv8 and input training set images for model training.

[0021] In this embodiment, the training set is established based on the public VisDrone2019-DET data set, and the training set images are input into the YOLO-LiteMax target detection model based on YOLOv8 of the present invention for model training.

[0022] Step 2: The input target object image is subjected to feature extraction through the backbone network to generate feature maps of different scales.

[0023] In step 201, the input image first passes through two CBS (Convolutions BnSiLU convolution) modules of the zeroth layer and the first layer, and is downsampled twice to obtain a 4-fold downsampled feature map, wherein the CBS module consists of a convolution kernel of size 3 and stride 2, a batch normalization layer, and a SiLU activation function.

[0024] Next, the 4x downsampled feature map is input into the SCB (Selective Convolution Block) module of the second layer for convolution.

[0025] In this embodiment, the SCB module is improved based on the traditional C2f module. The Bottleneck in the C2f module is replaced with a FasterNet block. The feature map is first processed by a CBS module in the SCB module, and then output to a Split layer. The SplitC2f module divides the feature map into two feature maps with the same number of channels. The output of each FasterNet block will be used as the input of the next FasterNet block and stored in a list. After passing through 3 FasterNet blocks, all feature maps in the list will be spliced ​​by channel.

[0026] It should be noted that in the original YOLOv8, a large number of Bottleneck structures bring high computational overhead, while this embodiment uses FasterNet blocks. Since each FasterNet block only convolves part of the channels, it achieves the purpose of reducing the model complexity and optimizing the performance of the deep neural network.

[0027] Next, the CBS module of the third layer performs downsampling again, that is, 8 times downsampling, and the 8 times downsampled feature map is input into the SCB module of the fourth layer for feature extraction.

[0028] Then, the fifth-layer CBS module performs downsampling again, that is, 16-fold downsampling, and the 16-fold downsampled feature map is input into the sixth-layer SCB module for feature extraction.

[0029] Then, the CBS module of the seventh layer performs downsampling again, that is, 32 times downsampling, and the 32 times downsampled feature map is input into the SCB module of the eighth layer for feature extraction.

[0030] Finally, the feature map P5 after 32 times downsampling passes through the ninth-layer SPPF module to improve the model's ability to capture contextual details.

[0031] In particular, the SPPF module consists of CBS and Maxpool operations to enhance feature extraction by pooling information at multiple scales.

[0032] Step 202, the feature maps output by the fourth, sixth, and eighth layer SCB modules in the Backbone of YOLOv8 are output to the neck network (Neck), that is, the 8-fold downsampling, 16-fold downsampling, and 32-fold downsampling feature maps after feature extraction by the SCB module; and the second layer of the SCB module is added, that is, the large-size feature map after 4-fold downsampling, and it is output to the Neck at the same time.

[0033] Step 3: Input the generated feature maps of different scales into the neck network and use the scale sequence for feature fusion.

[0034] Step 301, modify the connection mode between Backbone and Neck, from the original four, six, and nine layers connected to Neck to the second, fourth, sixth, and ninth layers connected to Neck; In particular, the feature map P5 output by the ninth-layer SPPF module is first convolved by a tenth-layer CBS module after entering the neck network, and the feature maps of the second, fourth, sixth, and fourteenth layers input to Neck are P2, P3, P4, and P5 feature maps, respectively.

[0035] In step 302, the feature maps of P3, P4, and P5 (8x downsampling, 16x downsampling, and 32x downsampling) are first input into the first SAC (Scale Aware Concatenation) module for feature fusion to enhance the ability to capture small target detail information.

[0036] The specific operation of the SAC module is: first take the large, medium and small size feature maps as input, perform adaptive maximum pooling and adaptive average pooling on the large size feature map and then add the two pooling results, perform interpolation upsampling operation on the small size feature map, and finally splice the processed large, medium and small size feature maps in the channel dimension.

[0037] The SAC module not only adds a large-scale feature layer, but also performs differentiated processing on feature maps of different scales, which not only balances the detail retention and global information extraction of feature maps, but also avoids the loss of small target features during upsampling.

[0038] Preferably, in this embodiment, the SAC feature fusion module formula is expressed as: ; in, Represents the result output by the SAC module. , and Represent large, medium and small size feature maps respectively, represents the maximum pooling operation, represents the average pooling operation, represents the interpolation upsampling operation, Represents a concatenation operation.

[0039] The feature map input by SAC is processed again by the SCB module, and then the P4-scale feature map output by SCB is input into the second SAC module together with the P2 and P3-scale feature maps to further fully fuse the features; then it is processed by the SCB module to obtain the P3-scale feature map, which is output to the P3-scale detection head for use by the P3 detection head.

[0040] Then, a downsampling operation is performed through a CBS module composed of a convolution kernel of size 3 and a stride of 2. The downsampling result is concat-operated with the feature map fused by the first SAC module, and processed by the SCB module to obtain a P4-scale feature map, which is output to the P4-scale detection head for use by the P4 detection head.

[0041] The above feature map is downsampled again through the CBS module composed of a convolution kernel of size 3 and step size 2. The downsampling result is concat-operated with the feature map output by SPPF, processed by the SCB module, and a P5-scale feature map is obtained, which is output to the P5-scale detection head for use by the P5 detection head.

[0042] Step 303: In order to fully extract feature map information, in this embodiment, two C3DSF (Convolution 3D Scale Fusion) feature extraction modules are added to Neck to capture the correlation between feature maps across scales. The input of the first C3DSF module is the feature maps P3, P4, and P5, and the input of the second C3DSF module is the feature maps P2, P3, and P4.

[0043] Specifically, the feature maps P3, P4, and P5 are used as inputs of the first C3DSF module for feature extraction.

[0044] The specific operation of feature extraction is to adjust the large, medium, and small-sized feature maps to the same number of channels, and then use interpolation upsampling operations on the medium and small-sized feature maps to adjust the feature map size to the size of the large-sized feature map, and then expand the dimensions of the large, medium, and small-sized feature map tensors and splice them. 3D convolution is used to extract features from the spliced ​​feature maps, and then they pass through the normalization layer, activation function layer, and maximum pooling layer, and finally the results of the above processing are compressed to the previous dimensions.

[0045] Compared with traditional 2D convolution, this embodiment performs more effective information fusion between feature layers of different scales, can capture the correlation between scales, and avoids information loss or discontinuity problems caused by simple splicing.

[0046] Preferably, in this embodiment, the C3DSF feature extraction module expression is: ; in, , , Represents a feature map of three sizes with an added 3D dimension, Represents a splicing operation, represents a three-dimensional convolution, represents the batch normalization operation, represents the Leaky ReLU activation function, represents the maximum pooling operation, Indicates dimensionality reduction operation, dim represents dimension, and dim=2 in the formula corresponds to the third dimension (index starts from 0).

[0047] Step 304: Add the features extracted by the first C3DSF module to the feature map fused by the SAC module.

[0048] Considering that adding the third SAC module will utilize the P1 layer feature map, which will greatly increase the parameters and the amount of calculation, this embodiment chooses to fuse the P2 layer through simple upsampling and Concat operations.

[0049] Next, an upsampling module and a Concat module are added to upsample the fused P3-scale features, and then Concat the output of the SCB module in the second layer of the backbone. After being processed again by the SCB module, the feature maps P2, P3, and P4 are first input into the second C3DSF module for feature extraction, and then added with the features processed by the above SCB to finally obtain the feature map of P2 scale, which is output to the P2-scale detection head for use by the P2 detection head.

[0050] Step 4: Input the feature map after the neck network features are fused into the head network (Head) for classification and regression tasks.

[0051] Step 401, modify the connection mode between Neck and Head from three layers of feature maps P3, P4, and P5 to four layers of feature maps P2, P3, P4, and P5, and add a feature map P2 (4 times downsampling) scale detection head.

[0052] Step 402, integrating group normalization into the CBS module. In this embodiment, BatchNormalization is replaced with GroupNormalization for improvement, and the improved module is named CGS.

[0053] Step 403: The feature map of the P2 scale is first subjected to an independent CGS (Conv, Group Normalization, SiLU) module for further feature extraction, and then subjected to two P2, P3, P4, and P5 scale feature maps. Figure 1 The CGS block with shared parameters is processed, and the features output by CGS are finally output to the P2, P3, P4, and P5 scale features. Figure 1The classification head and regression head of the shared 1×1 convolution are decoupled for learning; since sharing a single parameter will lead to inconsistent scales of the detected targets on the detection head, the regression features are scaled by the Scale factor, and finally Bbox_loss and Cls_loss are calculated separately.

[0054] Step 404: The feature map of the P3 scale is first subjected to an independent CGS module for further feature extraction, and then subjected to two CGS modules composed of the P2, P3, P4, and P5 scale features. Figure 1 The CGS block with shared parameters is processed, and the features output by CGS are finally output to the P2, P3, P4, and P5 scale features. Figure 1 The classification head and regression head of the shared 1×1 convolution are decoupled for learning; since sharing a single parameter will lead to inconsistent scales of the detected targets on the detection head, the regression features are scaled by the Scale factor, and finally Bbox_loss and Cls_loss are calculated separately.

[0055] Step 405: The feature map of the P4 scale is first subjected to an independent CGS module for further feature extraction, and then subjected to two CGS modules composed of the P2, P3, P4, and P5 scale features. Figure 1 The CGS block with shared parameters is processed, and the features output by CGS are finally output to the P2, P3, P4, and P5 scale features. Figure 1 The classification head and regression head of the shared 1×1 convolution are decoupled for learning; since sharing a single parameter will lead to inconsistent scales of the detected targets on the detection head, the regression features are scaled by the Scale factor, and finally Bbox_loss and Cls_loss are calculated separately.

[0056] Step 406: The feature map of the P5 scale is first subjected to an independent CGS module for further feature extraction, and then subjected to two CGS modules composed of the P2, P3, P4, and P5 scale features. Figure 1 The CGS block with shared parameters is processed, and the features output by CGS are finally output to the scale features P2, P3, P4, and P5 in Figure 1. Figure 1 The classification head and regression head of the shared 1×1 convolution are decoupled for learning; since sharing a single parameter will lead to inconsistent scales of the detected targets on the detection head, the regression features are scaled by the Scale factor, and finally Bbox_loss and Cls_loss are calculated separately.

[0057] The head network of this embodiment enhances the feature consistency and detection performance of the model in multi-scale target detection by applying a shared convolution module. It also enables detection heads of different scales to jointly optimize parameters during training through collaborative learning, effectively improving the robustness and generalization ability of feature extraction, thereby significantly improving the overall performance of the YOLO-LiteMax target detection network based on YOLOv8.

[0058] Step 5: Input the drone image to be detected into the trained YOLO-LiteMax target detection model to perform target detection and identify dense small targets in the image. The specific method is as follows:

[0059] Step 501, input the image into the backbone network of the model to extract a multi-scale feature map.

[0060] Step 502, through the small object scale sequence fusion network, the features of different scales are fused to enhance the model's detection capability for multi-scale objects.

[0061] Step 503, apply the prediction head on the feature maps of different scales to generate bounding boxes, class probabilities and confidence scores.

[0062] Step 504 , filtering the prediction boxes with lower confidence and suppressing the highly overlapping redundant boxes through post-processing.

[0063] Step 505 , draw the detected object bounding box, category label and confidence score on the original image to generate a visualization result.

[0064] Furthermore, the YOLO-LiteMax target detection network based on YOLOv8 of the present invention is compared with other YOLO series models, and the results are shown in Table 1 below: Table 1 Results of YOLO-LiteMax compared with other YOLO series models

[0065] It can be seen that compared with other YOLO series models, the YOLO-LiteMax target detection network of the present invention performs best in terms of precision, recall, average precision under 0.5 intersection-over-union ratio, and average precision mean in the range of 0.5 to 0.95 intersection-over-union ratio, while having the least number of parameters.

[0066] Furthermore, the YOLO-LiteMax target detection network based on YOLOv8 of the present invention is compared with the models of other classic neural network series, and the results are shown in Table 2 below: Table 2 Results of YOLO-LiteMax of the present invention compared with other classic models

[0067] It can be seen that compared with other classic models, the effect of the YOLO-LiteMax target detection network of the present invention is also far ahead.

[0068] In an embodiment of the present invention, an electronic device is also provided, including: one or more processors; a storage device on which one or more programs are stored; when the one or more programs are executed by the one or more processors, the one or more processors implement the drone target detection method based on the YOLO-LiteMax neural network described in any of the above embodiments.

[0069] In an embodiment of the present invention, a computer-readable storage medium is further provided, on which a computer program is stored. When the program is executed by a processor, the steps in any one of the drone target detection methods based on the YOLO-LiteMax neural network in the above embodiments are implemented.

[0070] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A drone target detection method based on YOLO-LiteMax neural network, characterized in that: The steps include: Step 1: Build a YOLO-LiteMax target detection model based on YOLOv8 and input training set images for model training; Step 2: The input image is subjected to feature extraction through the backbone network to generate feature maps of different scales; the backbone network includes several CBS modules, several selection convolution blocks, and an SPPF module; the selection convolution module replaces the Bottleneck in the C2f module with the FasterNet block to reduce redundant information in the feature map; Step 3: Input the generated feature maps of different scales into the neck network, and use the scale sequence to perform feature fusion; the neck network performs differential processing on the feature maps of different scales, fuses the local details and global information of the multi-scale feature maps, and then highlights the small target information by adding large-scale feature maps, and uses the feature extraction module to obtain the local details of the small target; Step 4: Input the feature map after the neck network features are fused into the head network for classification and regression tasks; in the head network, the detection head further extracts features using group normalization and shared convolution, shares information between features of different scales, and trains the classification and regression tasks separately through the decoupling head, and scales the feature map using the scale factor to obtain the trained YOLO-LiteMax target detection model; Step 5: Input the drone image to be detected into the trained YOLO-LiteMax target detection model to perform target detection and identify dense small targets in the image.

2. The UAV target detection method based on YOLO-LiteMax neural network according to claim 1, characterized in that: In step 2, each of the CBS modules consists of a convolution kernel of size 3 and stride 2, a batch normalization layer, and a SiLU activation function; Each of the selected convolution blocks includes a CBS module, a Split layer, and three sequentially connected FasterNet blocks; each FasterNet block performs convolution on only part of the channels, and the output of each FasterNet block is used as the input of the next FasterNet block and stored in a list. After three FasterNet blocks, all feature maps in the list are spliced ​​by channel; The SPPF module consists of CBS and Maxpool operations, which enhances the feature extraction capability of the backbone network by pooling information at multiple scales.

3. The UAV target detection method based on YOLO-LiteMax neural network according to claim 2 is characterized in that: In step 2, in the backbone network, the target object image is subjected to feature extraction through the backbone network as follows: Step 201: The input image is downsampled twice by two CBS modules in the zeroth layer and the first layer to obtain a 4-fold downsampled feature map, and enters the second layer to select the convolution block for convolution; In the second layer of convolutional blocks, the 4x downsampled feature map is first processed by a CBS module and the output enters a Split layer. The Split layer divides the feature map into two feature maps with the same number of channels, and then passes through three FasterNet blocks for feature extraction to obtain a large-size feature map P2 after 4x downsampling. The feature map P2 after 4 times downsampling is passed through the third layer CBS module, downsampled 8 times, and then input into the fourth layer selection convolution block for feature extraction to obtain the feature map P3 after 8 times downsampling; After passing through the fifth layer CBS module, it is downsampled 16 times and then input into the sixth layer selection convolution block for feature extraction to obtain the feature map P4 after 16 times downsampling; After passing through the seventh layer CBS module, it is downsampled 32 times and then input into the eighth layer selection convolution block for feature extraction to obtain the feature map P5 after 32 times downsampling; the feature map P5 is passed through the ninth layer SPPF module to improve the ability of the YOLO-LiteMax target detection network to capture context details; Step 202: Output the 8-fold downsampled feature map P3, the 16-fold downsampled feature map P4, the 32-fold downsampled feature map P5, and the 4-fold downsampled feature map P2 after feature extraction to the Neck network.

4. The UAV target detection method based on YOLO-LiteMax neural network according to claim 1, characterized in that: Step 3 In the neck network, feature fusion is performed on feature maps of different scales as follows: Step 301, modify the connection mode between the backbone network and the neck network, connect the fourth, sixth, and ninth layers of the traditional YOLOv8 backbone network to the neck network, and optimize the backbone network to connect the second, fourth, sixth, and ninth layers to the neck network; wherein, the feature map P5 output by the ninth-layer SPPF module is first convolved by the tenth-layer CBS module after entering the neck network; Step 302: The feature maps P3, P4 and the feature map P5 after convolution processing are used as feature maps of large, medium and small sizes, and are input into the first SAC module for feature fusion, and feature maps of different scales are differentiated to balance the detail retention and global information extraction of the feature maps, so as to avoid the loss of small target features during upsampling; The feature map output by the first SAC module is processed again by a selective convolution block, and the feature map output by the selective convolution block and the feature maps P2 and P3 are input to the second SAC module again to further fully fuse the features, and then processed by the selective convolution block to obtain the P3 scale feature map for use by the P3 detection head; The feature map output by the second SAC module is downsampled through a CBS module consisting of a convolution kernel of size 3 and a stride of 2. The downsampled result is concat-operated with the feature map fused by the first SAC module. After processing by selecting a convolution block, a P4 scale feature map is obtained for use by the P4 detection head. The obtained feature map is downsampled again through a CBS module consisting of a convolution kernel of size 3 and a stride of 2. The downsampled result is concat-operated with the feature map output by the tenth layer of the CBS module. After processing by the selected convolution block, the P5 scale feature map is obtained for use by the P5 detection head. Step 303: Add two C3DSF feature extraction modules to the neck network to capture the correlation between feature maps across scales, input feature maps P3, P4, and P5 into the first C3DSF feature extraction module to fully extract feature map information; Step 304, add the features extracted by the first C3DSF feature extraction module and the feature map fused by the second SAC module, add an upsampling module and a Concat module, fuse the P3 scale feature map by upsampling, and perform a Concat operation with the output of the second layer selected convolution block in the backbone network. After being processed again by the selected convolution block, the feature maps P2, P3, and P4 are input into the second C3DSF feature extraction module for feature extraction, and then add the features processed by the selected convolution block to obtain the P2 scale feature map for use by the P2 detection head.

5. The UAV target detection method based on YOLO-LiteMax neural network according to claim 4 is characterized in that: In the SAC module, adaptive maximum pooling and adaptive average pooling are performed on the large-size feature map, the results of the two poolings are added, an interpolation upsampling operation is performed on the small-size feature map, and then the processed large, medium, and small-size feature maps are spliced ​​in the channel dimension; The SAC module, the feature fusion expression is as follows: ； in, Represents the result output by the SAC module. , and Represent large, medium and small size feature maps respectively, represents the maximum pooling operation, represents the average pooling operation, represents the interpolation upsampling operation, Represents a concatenation operation.

6. The UAV target detection method based on YOLO-LiteMax neural network according to claim 4 is characterized in that: In the C3DSF feature extraction module, the large, medium and small size feature maps are adjusted to the same number of channels, the medium and small size feature maps are subjected to interpolation upsampling operations, the feature map size is adjusted to the large size feature map size, the dimensions of the large, medium and small size feature map tensors are expanded and spliced, 3D convolution is used to extract features from the spliced ​​feature maps, and then after passing through a normalization layer, an activation function layer and a maximum pooling layer, the obtained processing result is compressed to the dimension at the time of input; The C3DSF feature extraction module, the feature extraction expression is as follows: ； in, , , Represents a feature map of three sizes with an added 3D dimension, Represents a splicing operation, represents a three-dimensional convolution, represents the batch normalization operation, express Activation function, represents the maximum pooling operation, represents the dimensionality reduction operation, Represents dimension.

7. The UAV target detection method based on YOLO-LiteMax neural network according to claim 4 is characterized in that: In step 4, in the head network, classification and regression tasks are performed as follows: Step 401, modify the connection mode between the neck network and the head network, add a P2 scale detection head, connect the three-layer scale feature map of the traditional YOLOv8 neck network P3, P4, and P5 to the head network, and modify it to connect the four-layer scale feature map of the neck network P2, P3, P4, and P5 to the head network; Step 402: Integrate group normalization into the head network CBS module, replace BatchNormalization with GroupNormalization, and obtain an improved CGS module; Step 403, the P2 scale feature map is first subjected to further feature extraction by an independent CGS module, and then processed by two CGS modules whose parameters are shared by the P2, P3, P4, and P5 scale feature maps. The features output by the CGS module are respectively output to the classification head and regression head of the 1×1 convolution shared by P2, P3, P4, and P5 for decoupled learning, and the regression features are scaled by the Scale factor, and the Bbox_loss and Cls_loss of the P2 scale feature map are respectively calculated; Step 404, the P3 scale feature map is first subjected to further feature extraction by an independent CGS module, and then processed by two CGS modules whose parameters are shared by the P2, P3, P4, and P5 scale feature maps. The features output by the CGS module are respectively output to the classification head and regression head of the 1×1 convolution shared by P2, P3, P4, and P5 for decoupled learning, and the regression features are scaled by the Scale factor, and the Bbox_loss and Cls_loss of the P3 scale feature map are respectively calculated; Step 405, the P4 scale feature map is first subjected to further feature extraction by an independent CGS module, and then processed by two CGS modules whose parameters are shared by the P2, P3, P4, and P5 scale feature maps. The features output by the CGS module are respectively output to the classification head and regression head of the 1×1 convolution shared by P2, P3, P4, and P5 for decoupled learning, and the regression features are scaled by the Scale factor, and the Bbox_loss and Cls_loss of the P4 scale feature map are respectively calculated; Step 406, the P5 scale feature map is first further extracted by an independent CGS module, and then processed by two CGS modules whose parameters are shared by the P2, P3, P4, and P5 scale feature maps. The features output by the CGS module are respectively output to the classification head and regression head of the 1×1 convolution shared by P2, P3, P4, and P5 for decoupled learning, and the regression features are scaled by the Scale factor, and the Bbox_loss and Cls_loss of the P5 scale feature map are calculated respectively.

8. The UAV target detection method based on YOLO-LiteMax neural network according to claim 7, characterized in that: The Bbox_loss and Cls_loss of the obtained P2, P3, P4, and P5 scale feature maps are used. The regression loss Bbox_loss is used to optimize the position and size of the predicted bounding box to ensure that the object is accurately located; the classification loss Cls_loss is used to improve the accuracy of category recognition to ensure that the object category is correctly classified. By minimizing these two losses at the same time, the YOLO-LiteMax target detection network achieves a balance between positioning and classification. By building a CGS module in the head network and applying shared convolution, the feature consistency and detection performance of the YOLO-LiteMax target detection network in multi-scale target detection are enhanced. Through collaborative learning, the detection heads of different scales are jointly optimized during training to improve the robustness and generalization ability of feature extraction.

9. An electronic device, characterized in that: include: one or more processors; a storage device having one or more programs stored thereon; When the one or more programs are executed by the one or more processors, the one or more processors implement the drone target detection method based on the YOLO-LiteMax neural network as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the program is executed by a processor, the steps in the drone target detection method based on the YOLO-LiteMax neural network described in any one of claims 1 to 8 are implemented.

Images