Lightweight target detection method for automatic artillery filling system based on YOLO-AGU

Abstract

The invention discloses a lightweight target detection method for an automatic artillery loading system based on YOLO-AGU, and the method improves a YOLOv11 architecture, and comprises the steps: embedding an AOD-Net image enhancement module at the front end of a backbone network, and improving the image quality through multi-scale convolution feature extraction; a GSConv module is combined with the advantages of standard convolution and depth separable convolution, and channel shuffling operation is matched to enhance the feature expression ability; and constructing an efficient feature fusion architecture containing a UI B module, and optimizing a gradient information transmission path. According to the method, through an innovative network structure design, the detection performance in a low-light environment is remarkably improved while the light weight is kept, and the method is suitable for a real-time target detection task of the rotary missile cabin of the automatic artillery filling system.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and target detection technology, specifically relating to a lightweight target detection method for an automatic artillery loading system based on YOLO-AGU, which is particularly suitable for real-time target detection tasks in low-light environments of automatic artillery loading systems. Background Technology

[0002] As a core component of modern artillery equipment, the reliable operation of automatic artillery loading systems directly impacts combat effectiveness. Current computer vision-based automated detection technologies face severe challenges in practical applications: First, the harsh environment of the loading compartment, characterized by smoke and complex lighting, significantly degrades the performance of traditional detection algorithms; the standard YOLOv11 model achieves only 77.2% mAP@50-95 in this environment. Second, existing image enhancement algorithms, such as those based on Retinex theory, generally require over 15 GFLOPs of computation, making them unsuitable for real-time performance. Third, complex network structures like Faster R-CNN have excessively large parameter counts, hindering effective deployment on vehicle-mounted embedded devices. Therefore, developing a target detection method that can effectively improve detection performance in harsh environments while maintaining a lightweight design has become a key technological requirement for solving the monitoring challenges of automatic artillery loading systems. Summary of the Invention

[0003] This invention proposes a lightweight target detection method for an automatic artillery loading system based on YOLO-AGU. This method, based on the baseline model YOLOv11, designs a YOLO-AGU model to adapt to high-noise environments and optimize computational efficiency. The model integrates the ultra-lightweight image denoising module AOD-Net and optimizes feature fusion through the PONO architecture. Simultaneously, to address computational resource limitations, the C3k2 module is optimized using the UIB architecture and paired with GSConv convolutions to improve feature fusion efficiency. The lightweight GUNeck, composed of UIB-C3k2 and GSConv, replaces the original neck network, further reducing redundant parameters and improving model performance. To enhance bounding box localization accuracy, MPDIoU is introduced to optimize the loss function and reduce rotation-based quantization analysis errors. Compared to YOLOv11, YOLO-AGU significantly improves detection accuracy, speed, and resource efficiency, meeting mission requirements.

[0004] The technical solution for achieving the present invention is: a lightweight target detection method for an automatic artillery loading system based on YOLO-AGU, characterized by the following steps:

[0005] Step S1: Use an action camera to capture video of the magazine rotation from multiple angles, perform frame rate extraction on the video to obtain an image dataset, and use the LabelImg annotation tool to annotate the images in the image dataset to obtain the ammunition loading system magazine dataset, i.e., the VOC dataset, and proceed to step S2.

[0006] Step S2: Filter the VOC dataset by category, retain the images and labels of projectiles, ammunition magazines and robotic arms, and convert the label format to TXT to obtain the filtered VOC dataset, then proceed to step S3.

[0007] Step S3: Simulate harsh environments by adding salt and pepper noise and adjusting exposure, perform data augmentation on the filtered VOC dataset to obtain a fused dataset, and randomly divide the fused dataset into a training set and a validation set at a ratio of 9:1, then proceed to step S4.

[0008] Step S4: Design the improved YOLO-AGU network, then proceed to step S5.

[0009] Step S5: Feed the training set into the improved YOLO-AGU network for training to obtain the YOLO-AGU model, and then proceed to step S6.

[0010] Step S6: Optimize the YOLO-AGU model using the MPDIoU gradient gain allocation strategy to obtain the optimized YOLO-AGU model, and then proceed to step S7.

[0011] Step S7: Input the validation set into the optimized YOLO-AGU model to obtain the optimized YOLO-AGU model results and the model's accuracy, thus achieving target detection.

[0012] Compared with the prior art, the significant advantages of this invention are:

[0013] (1) In this invention, AOD-Net captures different details of the image by using convolution kernels of different sizes and uses feature splicing to enhance the expressive power of the model, significantly improving the visibility and clarity of the input image in complex environments, ensuring that YOLOv11 can more accurately obtain target feature information in harsh environmental scenarios, and reducing the situation of missed detection and false detection.

[0014] (2) This invention uses the UIB-C2f and GSConv modules as the core, and combines the upsampling and feature concatenation modules to design and construct an efficient neural network neck structure GU-Neck, which maximizes the use of computing resources to reduce redundant parameters of the model while ensuring the recognition performance of the model.

[0015] (3) The present invention adjusts the loss function of the YOLO-AGU model and uses the MPDIoU loss function to minimize the distance between the upper left and lower right corners of the predicted bounding box and the true bounding box, thereby optimizing the model in a targeted manner to make it more accurate in target localization and recognition.

[0016] (4) This invention proposes a lightweight target detection method for an automatic artillery loading system based on YOLO-AGU. The YOLO-AGU model has improved in all aspects, especially in terms of accuracy. The optimized GU-Neck increases computational efficiency and reduces the occupation of computational resources, so it does not have a significant impact on the amount of computation and achieves a balance between accuracy and speed. Attached Figure Description

[0017] Figure 1 This is a flowchart of a lightweight target detection method for an automatic artillery loading system based on YOLO-AGU, according to the present invention.

[0018] Figure 2 This is a schematic diagram of the YOLO-AGU target detection based on the present invention.

[0019] Figure 3 This is a schematic diagram of AOD-Net in this invention.

[0020] Figure 4 This is a flowchart illustrating the instantiation process of UIB-C3k2 in this invention.

[0021] Figure 5 This is a diagram of the GSConv module in this invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0023] The technical solutions of the various embodiments of the present invention can be combined with each other, but only if they can be implemented by those skilled in the art. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0024] The following section will further introduce the specific implementation method, as well as the technical difficulties and inventive points of this invention, using this design example as an example.

[0025] Combination Figures 1-5The present invention discloses a lightweight target detection method for an automatic artillery loading system based on YOLO-AGU, comprising the following steps:

[0026] Step S1: Use an action camera to capture video of the magazine rotation from multiple angles, perform frame rate extraction on the video to obtain an image dataset, and use the LabelImg annotation tool to annotate the images in the image dataset to obtain the ammunition loading system magazine dataset, i.e., the VOC dataset, and proceed to step S2.

[0027] Step S2: Filter the VOC dataset by category, retain the images and labels of projectiles, ammunition magazines and robotic arms, and convert the label format to TXT to obtain the filtered VOC dataset, then proceed to step S3.

[0028] Step S3: Simulate harsh environments by adding salt and pepper noise and adjusting exposure, perform data augmentation on the filtered VOC dataset to obtain a fused dataset, and randomly divide the fused dataset into training and validation sets in a 9:1 ratio, then proceed to step S4.

[0029] Step S4: Design the improved YOLO-AGU model:

[0030] The overall architecture of the YOLO-AGU model is as follows: Figure 2 As shown. The improved YOLO-AGU model includes AOD-Net, a backbone network, a neck network, and a detection head. AOD-Net serves as the input layer and includes a K estimation module and a sharp image generation module. Proceed to step S5.

[0031] Step S5: Feed the training set into the improved YOLO-AGU model for training to obtain the trained YOLO-AGU model, as follows:

[0032] AOD-Net performs preprocessing and enhancement operations on the training set. Specifically, AOD-Net utilizes multi-scale convolution to fuse image features, improving the visibility and clarity of images in complex environments, resulting in several feature maps, as detailed below:

[0033] Let I(x) represent the images in the training set, i.e., the original images.

[0034] The K estimation module infers the variable K(x) based on the input original image I(x), and the sharp image generation module uses K(x) as an adaptive parameter to generate the feature map J(x);

[0035] J(x)=K(x)I(x)-K(x)+b,

[0036] Where I(x) represents the image in the training set, i.e. the original image, and the constant bias b = 1.

[0037] The variable K(x) is:

[0038]

[0039] Wherein, the variable t(x) is:

[0040] t(x)=e -βd(x) ,

[0041] Where β is the atmospheric scattering coefficient, d(x) is the distance between the object and the moving camera, and Z represents the atmospheric light value at infinity.

[0042] The core concept of AOD-Net lies in unifying the parameters Z and t(x) into a single formula, thereby directly minimizing pixel-level reconstruction errors. Z and are merged into a new variable K(x).

[0043] AOD-Net captures different details of images by using convolutional kernels of different sizes and uses feature concatenation to enhance the model's expressive power, significantly improving the visibility and clarity of input images in complex environments. This ensures that YOLOv11 can more accurately acquire target feature information in harsh environmental scenarios, reducing missed detections and false detections.

[0044] The aforementioned feature maps are then fed into the backbone network, as follows:

[0045] The backbone network employs the YOLO network to extract multi-scale feature information from the feature maps. To improve the efficiency and quality of feature extraction, UIB-C3k2 is introduced. UIB-C3k2 integrates repeated convolutional structures and an efficient information transfer mechanism, resulting in a more accurate network model. This significantly reduces the computational burden while enhancing feature representation capabilities. Simultaneously, the newly added C2PSA module possesses excellent information filtering capabilities, effectively suppressing redundant interference and guiding the model to focus on key regions in the image, thereby improving its target detection performance in complex scenes.

[0046] The neck network uses GU-Neck, which is an improvement on the YOLO network in the neck area. GU-Neck replaces the original neck network in YOLO. GU-Neck is based on the UIB-C2f and GSConv modules, combined with upsampling and feature concatenation modules. This maximizes computational resource utilization, reduces redundant parameters, and ensures the model's recognition performance. UIB-C3k2 is an improvement that introduces the UIB mechanism into the C3k2 module, as detailed below:

[0047] UIB is built on top of MobileNetV4, employing a first depthwise separable convolution (DWConv), pointwise expansion (PWConv), and projection in the IB. Two optional second depthwise separable convolutions (DWConv) are introduced in the inverse bottleneck block (IB): one DWConv before the expansion layer and the other between the expansion and projection layers.

[0048] The Extra DW scheme was selected to optimize the neck network, which not only enhanced the network's feature extraction capabilities but also provided better robustness in handling complex scenes and low-light environments, compensating for the deficiency of C3k2 in losing certain details when processing complex images. GSConv was introduced to replace the standard convolution in the neck network to reduce the model's burden while maintaining accuracy, as detailed below:

[0049] Assume the size of the input multi-scale feature map is The target output dimension is H represents the height of the input channel. out W represents the height of the output channel, and W represents the width of the input channel. out C represents the width of the output channel. in Indicates the number of input channels, C out This represents the number of output channels, where C in =C out Established.

[0050] The calculation process for GSConv is as follows:

[0051] The input multi-scale feature map is divided into two parts: one part is... Perform a depthwise convolution on it to obtain a depthwise convolutional image; the other part is... Preserve the original information.

[0052] Perform a depthwise convolution on X1, assuming the kernel size is K1×K1 and the output dimension remains unchanged. The computational cost F of the depthwise convolution DWConv part is... DW for:

[0053]

[0054] The depthwise convolutional image is concatenated with S, and a shuffle operation involving only inter-channel rearrangement is performed to enhance information exchange. A 1×1 point convolution is then applied to the concatenated result to fuse information, yielding a fused feature map. The number of output channels for the point convolution part is C. out Its computational complexity F PW as follows:

[0055] F PW =H×W×Cin ×C out ,

[0056] Integration shows that the total computational cost of GSConv and parameter quantity Params GSConv for:

[0057]

[0058] The YOLO-AGU model has improved across all metrics, especially in terms of accuracy, which has increased significantly. Furthermore, the optimized GU-Neck has increased computational efficiency and reduced the consumption of computing resources, so it has not had a significant impact on the amount of computation, achieving a balance between accuracy and speed.

[0059] The feature map is finally fed into the detection head, and after detection, the predicted features are output to obtain the trained YOLO-AG U model, and then proceed to step S6.

[0060] Step S6: Employ the MPDIoU gradient gain allocation strategy. This metric, in addition to considering the overlap between the target and the prediction box, further incorporates the distance relationship between targets, resulting in the optimized YOLO-AGU model, as detailed below:

[0061] By introducing the loss function MPDIoU, which directly minimizes the distance between the top-left and bottom-right corners of the predicted bounding box and the ground truth bounding box, the degradation phenomenon of the loss function is avoided. The model is specifically optimized to make it more accurate in target localization and recognition, thus significantly improving the model's localization accuracy and convergence speed. The optimized YOLO-AGU model is as follows:

[0062]

[0063]

[0064] In this image, A and B are two rectangles, each uniquely identified by the coordinates of its top-left and bottom-right corners. This represents the coordinates of the top-left and bottom-right corners of rectangle A. This represents the coordinates of the top-left and bottom-right corners of rectangle B; w and h represent the width and height of the image, respectively; d1 represents... and The straight-line distance, d2 represents and The straight-line distance;

[0065] It is a classic IoU metric, incorporating the distance between the top-left and bottom-right corners into the IoU calculation to measure the degree of overlap between rectangles A and B; and As a penalty factor, MPDIoU is deducted from the calculation and divided by (w 2 +h 2 The purpose is to normalize the image size to ensure that the distance penalty is consistent across images of different sizes.

[0066] According to the definition of MPDIoU, the loss function is based on the definition of MPDIoU. The formula is as follows:

[0067]

[0068] Proceed to step S7.

[0069] Step S7: Input the validation set into the optimized YOLO-AGU model to obtain the optimized YOLO-AGU model results and the model's accuracy, thus achieving target detection.

[0070] Example 1

[0071] In this invention, the computing platform used in the experiment ran the Ubuntu 20.04 operating system and was equipped with Python 3.8. The RKNN Toolkit2 is compatible with multiple deep learning frameworks, including PyTorch, TensorFlow, ONNX, Darknet, and Caffe. Additionally, the experimental parameters were set to batch size of 64, worker count of 10, epoch count of 200, initial learning rate of 0.01, and network input size of 640*640.

[0072] Precision (P), recall (R), mean average precision (mAP), and frames per second (FPS) are used to evaluate the model's accuracy and detection speed.

[0073]

[0074]

[0075] Where TP (True Positive) represents the number of samples that the model predicts are positive and are actually positive, FP (False Positive) represents the number of samples that the model predicts are positive but are actually negative, and FN (False Negative) represents the number of samples that the model predicts are negative but are actually positive.

[0076] AP (Average Precision) is the area under the Recall and Precision curves, and mAP (Mean Average Precision) is the average of the average precision across all categories, reflecting the average recognition accuracy across all categories. The formulas for AP and mAP are as follows.

[0077]

[0078] Where C is the total number of categories, N is the number of samples in the test set, P(k) is the precision when k samples are identified simultaneously, and ΔR(k) is the change in recall when the number of detected samples changes from k-1 to k.

[0079] mAP@0.35 represents the mAP value with an IoU threshold of 0.5. mAP@[0.5:0.95] calculates the mAP for multiple thresholds ranging from 0.5 to 0.95 with a step size of 0.05, and then takes the average. The larger mAP@[0.5:0.95] is, the more accurate the model's predicted bounding box is.

[0080]

[0081] The balanced F-score, F1_score, is the harmonic mean of precision and recall, used to achieve a balance between the two.

[0082]

[0083] The YOLO-AGU model proposed in this invention achieves the following accuracy metrics: precision of 94.16%, an improvement of 3.35% compared to YOLOv11n and 6.71% compared to YOLOv8n; recall of 93.42%, an improvement of 7.28% compared to YOLOv11n and 7.21% compared to YOLOv8n; mAP@0.5 of YOLO-AGU is 93.26%, an improvement of 1.55% compared to YOLOv11n and 3.14% compared to YOLOv8n; and mAP@0.5-0.95 of YOLO-AGU is 82.79%, an improvement of 5.58% compared to YOLOv11n and 10.36% compared to YOLOv8n. The F1 score of YOLO-AGU is 93.79%, which is an improvement of 5.38% compared to YOLOv11n and 6.96% compared to YOLOv8n. Table 1 demonstrates that the YOLO-AGU model has improved across all metrics, especially in terms of accuracy, where it has seen a significant increase.

[0084] Table 1 Comparison of YOLO-AGU Model Parameters

[0085] Model Precision Recall mPA@0.5 mPA@0.5-0.95 F1 YOLOv8n 0.8745 0.8621 0.9012 0.7243 0.8683 YOLOv11n 0.9081 0.8614 0.9171 0.7721 0.8841 YOLO-AGU 0.9416 0.9342 0.9326 0.8279 0.9379

[0086] The scope of protection of this invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this invention should be included within the scope of protection of this invention. Therefore, the scope of protection of this invention should be determined by the scope of the claims.

Claims

1. A lightweight target detection method for an automatic artillery loading system based on YOLO-AGU, characterized in that, The steps are as follows: Step S1: Use an action camera to capture video of the magazine rotation from multiple angles, perform frame rate extraction on the video to obtain an image dataset, and use the LabelImg annotation tool to annotate the images in the image dataset to obtain the ammunition loading system magazine dataset, i.e., VOC dataset, and proceed to step S2. Step S2: Filter the VOC dataset by category, retain the images and label data of projectiles, ammunition magazines and robotic arms, and convert the label format to TXT to obtain the filtered VOC dataset, then proceed to step S3; Step S3: Simulate harsh environments by adding salt and pepper noise and adjusting exposure, perform data augmentation on the filtered VOC dataset to obtain a fused dataset, and randomly divide the fused dataset into a training set and a validation set in a 9:1 ratio, then proceed to step S4. Step S4: Design the improved YOLO-AGU network, then proceed to step S5; Step S5: Feed the training set into the improved YOLO-AGU network for training to obtain the YOLO-AGU model, and then proceed to step S6. Step S6: Optimize the YOLO-AGU model using the MPDIoU gradient gain allocation strategy to obtain the optimized YOLO-AGU model, and then proceed to step S7. Step S7: Input the validation set into the optimized YOLO-AGU model to obtain the optimized YOLO-AGU model results and the model's accuracy, thus achieving target detection.

2. The lightweight target detection method for an automatic artillery loading system based on YOLO-AGU according to claim 1, characterized in that: In step S4, an improved YOLO-AGU model is designed, including AOD-Net, backbone network, neck network, and detection head. AOD-Net is used as the input layer and includes a K estimation module and a sharp image generation module.

3. The lightweight target detection method for an automatic artillery loading system based on YOLO-AGU according to claim 2, characterized in that, In step S5, the training set is fed into the improved YOLO-AGU network for training to obtain the YOLO-AGU model, as follows: AOD-Net performs preprocessing and enhancement operations on the training set to obtain several feature maps. These feature maps are then fed into the backbone network, which extracts multi-scale feature information at different levels to obtain multi-scale feature maps. Subsequently, the neck network fuses and enhances the multi-scale feature maps to obtain fused feature maps. These fused feature maps are then fed into the detection head, and after detection, predicted features are output to obtain the YOL O-AGU model.

4. The lightweight target detection method for an automatic artillery loading system based on YOLO-AGU according to claim 3, characterized in that, AOD-Net performs preprocessing and enhancement operations on the training set. Specifically, AOD-Net utilizes multi-scale convolution to fuse image features, improving the visibility and clarity of images in complex environments, resulting in several feature maps, as detailed below: Let I(x) represent the images in the training set, i.e., the original images; The K estimation module infers the variable K(x) based on the input original image I(x), and the sharp image generation module uses K(x) as an adaptive parameter to generate the feature map J(x); J(x)=K(x)I(x)-K(x)+b, Where I(x) represents the images in the training set, i.e., the original images, and the constant bias b = 1; The variable K(x) is: Wherein, the variable t(x) is: t(x)=e -βd(x) , Where β is the atmospheric scattering coefficient, d(x) is the distance between the object and the moving camera, and Z represents the atmospheric light value at infinity.

5. The lightweight target detection method for an automatic artillery loading system based on YOLO-AGU according to claim 4, characterized in that: The aforementioned feature maps are then fed into the backbone network, as follows: The backbone network adopts the YOLO network to extract multi-scale feature information from the feature map. To improve the efficiency and quality of feature extraction, UIB-C3k2 is introduced. UIB-C3k2 integrates repeated convolutional structures and efficient information transmission mechanisms, thereby obtaining a network model with higher accuracy.

6. The lightweight target detection method for an automatic artillery loading system based on YOLO-AGU according to claim 5, characterized in that: The neck network uses GU-Neck, which is an improvement on the YOLO network in the neck area. The design uses GU-Neck to replace the original neck network in YOLO. GU-Neck includes UIB-C3k2, GSConv and Concat modules.

7. The lightweight target detection method for an automatic artillery loading system based on YOLO-AGU according to claim 6, characterized in that: UIB-C3k2 refers to the improvement made by introducing the UIB mechanism into the C3k2 module, as detailed below: UIB is built on top of MobileNetV4, which uses the first DWConv, PWConv and the projection IB; two optional second DWConv are introduced in the IB, one second DWConv before the extension layer and the other between the extension layer and the projection layer.

8. The lightweight target detection method for an automatic artillery loading system based on YOLO-AGU according to claim 7, characterized in that: The neck network performs feature fusion and enhancement on multi-scale feature maps to obtain a fused feature map. Specifically, the ExtraDW scheme is used to optimize the neck network, introducing GSConv to replace the standard convolutions in the neck network to reduce the model's burden while maintaining accuracy. The details are as follows: Assume the size of the input multi-scale feature map is The target output dimension is H represents the height of the input channel. out W represents the height of the output channel, and W represents the width of the input channel. out C represents the width of the output channel. in Indicates the number of input channels, C out This represents the number of output channels, where C in =C out Established; The input multi-scale feature map is divided into two parts: one part is... Perform a depthwise convolution on it to obtain a depthwise convolutional image; the other part is... Preserve the original information; The deep convolutional image is concatenated with X2, and a shuffle operation involving only inter-channel rearrangement is performed to enhance information interaction. The concatenated result is then subjected to a 1×1 point convolution to fuse information and obtain a fused feature map.

9. The lightweight target detection method for an automatic artillery loading system based on YOLO-AGU according to claim 8, characterized in that: In step S6, the YOLO-AGU model is optimized using the MPDIoU gradient gain allocation strategy to obtain the optimized YOLO-AGU model, as follows: By introducing the loss function MPDIoU, which directly minimizes the distances between the top-left and bottom-right corners of the predicted bounding box and the ground truth bounding box, the degradation phenomenon of the loss function is avoided, thus significantly improving the localization accuracy and convergence speed of the model. The optimized YOLO-AGU model is as follows: MPDIoU makes the model focus more on the spatial distribution patterns and structural consistency of targets in the dataset. Its formula is: In this image, A and B are two rectangles, each uniquely identified by the coordinates of its top-left and bottom-right corners. This represents the coordinates of the top-left and bottom-right corners of rectangle A. This represents the coordinates of the top-left and bottom-right corners of rectangle B; w and h represent the width and height of the image, respectively; d1 represents... and The straight-line distance, d2 represents and The straight-line distance; It is a classic IoU metric, incorporating the distance between the top-left and bottom-right corners into the IoU calculation to measure the degree of overlap between rectangles A and B; and As a penalty factor, MPDIoU is deducted from the calculation and divided by (w 2 +h 2 This is to normalize the image size to ensure that the distance penalty is consistent across images of different sizes; According to the definition of MPDIoU, the loss function is based on the definition of MPDIoU. The formula is as follows:

Images