Low-altitude unmanned aerial vehicle target detection optimization algorithm based on deep learning

Through DBSCAN and K-means clustering optimization, Anchor Box is combined with pyramid feature fusion and DIoU-NMS, the problem of insufficient target detection accuracy and real-time performance of low-altitude drones in the existing technology is solved, and efficient and accurate drones are achieved. Target recognition.

CN119048786AInactive Publication Date: 2024-11-29LOONGRISE AVIONICS CO LTD
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202411535090.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-10-31
Publication Date
2024-11-29
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The existing low-altitude drone object detection algorithm based on deep learning has shortcomings in small object detection accuracy and real-time performance, especially the poor performance of image underlying feature details capture and detection in complex environments.

Method used

The Anchor Box scale is optimized by DBSCAN and K-means clustering algorithm, and the pyramid feature fusion module and multi-branch residual unit are introduced, combined with sub-pixel convolution, and feature fusion capability is improved. The traditional NMS is replaced by the DIoU-NMS algorithm to optimize the detection box assessment and inhibition.

Benefits of technology

It significantly improves the accuracy and effectiveness of drone target detection, enhances detection performance in complex environments, reduces missed detection problems, and ensures the integrity and detection accuracy of image semantic information.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119048786A_ABST
    Figure CN119048786A_ABST
Patent Text Reader

Abstract

The invention discloses a low-altitude unmanned aerial vehicle target detection optimization algorithm based on deep learning, and the algorithm improves the SSD algorithm, employs a DBSCAN + K-Means clustering algorithm to optimize the scale of an Anchor Box, improves the matching probability of a target object and the Anchor Box, combines a multi-branch residual unit, sub-pixel convolution, and a feature pyramid to obtain the multi-scale fusion feature of an image, and achieves the target detection of a low-altitude unmanned aerial vehicle. And finally, a DIOU-NMS algorithm is introduced to replace a traditional NMS algorithm to process the problem of a redundant frame in the image target, so that the missing detection of the target is further reduced, and the target detection effect of the low-altitude unmanned aerial vehicle is further improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of UAV algorithms, and in particular to a low-altitude UAV target detection optimization algorithm based on deep learning. Background Art

[0002] With the rapid development of deep learning, object detection has made significant progress. Object detection algorithms based on convolutional neural networks are divided into two categories: single-stage object detection algorithms and two-stage object detection algorithms. The single-stage object detection algorithm can locate objects through regular and dense sampling grids on the image, realizing end-to-end object detection. Compared with the two-stage object detection algorithm, it has higher detection accuracy and detection speed, such as YOLO, SSD, etc. Among them, the SSD algorithm, as an advanced object detection technology, can directly predict the category of the object and the position of its bounding box. It demonstrates high adaptability and robustness to changes in target scale by deeply integrating multiple convolutional layer architectures.

[0003] In order to improve the accuracy of small target detection, based on the SSD algorithm, DSSD uses a more complex ResNet-101 to replace the VGG16 network foundation to enhance feature learning capabilities, but with the expansion of parameter scale and the increase in computational complexity, the real-time performance of detection is affected; FPN-SSD introduces a top-down semantic enhancement path to enhance the richness of shallow features by fusing adjacent feature maps, but ignores the influence of low-level features on high-level features, resulting in insufficient feature fusion. The RSSD model combines pooling technology with deconvolution to achieve cross-scale feature fusion, thereby improving detection accuracy while maintaining a certain detection speed, but it still has shortcomings in focusing on small target feature extraction, especially capturing details of the underlying features of the image; the DF-SSD model is based on the innovation of dense convolution and function fusion, and shows good performance in detecting small targets and targets with specific spatial relationships. However, this performance improvement comes at the cost of a more complex network structure. Although it has achieved higher detection accuracy, it is still difficult to meet the needs of real-time detection.

[0004] In summary, a low-altitude UAV target detection optimization algorithm based on deep learning is needed to solve the shortcomings of the existing technology. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention provides a low-altitude UAV target detection optimization algorithm based on deep learning, aiming to solve the above problems.

[0006] To achieve the above object, the present invention provides the following technical solution: a low-altitude UAV target detection optimization algorithm based on deep learning, comprising the following steps: Step S1: Train the SSD target detection model using a publicly labeled data set to obtain an initial SSD target detection model; Step S2: Calculate the optimal aspect ratio of the Anchor Box according to the Anchor Box scale optimization strategy and generate the Anchor Box; Step S3: Introduce the pyramid feature fusion module to fuse the deep features and the shallow feature map semantic information; Step S4: Calculate the loss value by using the SSD loss function; Step S5: Perform back propagation according to the loss value and update the model gradient; Step S6: Use the DIoU-NMS algorithm to eliminate overlapping detection frames and obtain the detection frame with the highest confidence, that is, the optimal detection result.

[0007] Optionally, the model training in step S1 is: Step S11: Obtain a publicly labeled data set and divide it into a training set, a validation set, and a test set; Step S12: Use the VGG16 convolutional neural network as the basic network to train the SSD target detection model to obtain the SSD target detection initial model.

[0008] Optionally, the distance formula of the K-means algorithm in the clustering process is: , Where B is the sample box size; C is the cluster center generated by the K-means clustering algorithm; A is the cluster box size; IoU is the interaction ratio of the sample box size and the cluster box size, which is used as a measure of the similarity of the two box sizes.

[0009] Optionally, the Anchor Box is calculated and generated in step S2 in the following manner: Step S21: Data preprocessing: extract the size of each image and the corresponding target frame information from the training data set, and calculate the width and height of the target frame; Step S22: Image and target frame scaling: the image and its target frame are adaptively scaled proportionally so that the image and the target frame are analyzed at the same size; Step S23: Initialize Anchor Box: randomly select K sample points in the scaled target box size set as the initial clustering centers; Step S24: Distance calculation and classification: Calculate the distance between each target box and all cluster centers, and assign each target box to the nearest Anchor Box; Step S25: removing noise points by using DBSCAN clustering algorithm to optimize the location of cluster centers; Step S26: Repeat steps S24 to S25 until the cluster center position is stable and the final Anchor Box is generated.

[0010] Optionally, step S3 is implemented in the following manner: Step S31: introducing a multi-branch residual unit to fuse image features from different receptive fields; Step S32: Sub-pixel convolution improves the spatial resolution of the image by rearranging the pixel values ​​in the feature map channel through a specific convolution operation, and outputs a high-resolution image with a width and height that are r times that of the original image.

[0011] Optionally, the step S4 is calculated in the following manner: The target loss function is divided into two parts: the confidence loss of the predicted box and the target category, and the loss of the corresponding position regression.

[0012] Optionally, the calculation formula of the objective loss function is: , Where N is the number of matches between the predicted box and the real box, x represents the matching result, c and l represent the category confidence and location information of the predicted result respectively, g represents the real box information, and the parameter α is used to adjust the ratio between the location loss and the confidence loss, which is set to 1, L conf Represents the confidence loss, specifically using Softmax Loss as the loss function, L loc Represents the position regression loss, and specifically uses SmoothL1 Loss as the loss function.

[0013] Optionally, the model gradient is updated in step S5 as follows: By calculating the loss value obtained by the loss function, starting from the output layer, the partial derivative of each neuron is calculated reversely layer by layer according to the chain rule of the error, the adjustment amount of each weight is calculated according to the stochastic gradient descent rule, and the adjusted weights are applied to the network to update the value.

[0014] Optionally, the calculation formula of the DIoU-NMS algorithm in step S6 is: , Where si is the classification confidence; ε is the NMS threshold, and IoU is the intersection-over-union ratio of the sample box size and the clustering box size. is the distance between the center points of box b and box bgt; c represents the length of the diagonal of the smallest rectangle that can cover both boxes at the same time.

[0015] Beneficial effects of the present invention: 1. The present invention combines the advantages of DBSCAN and K-means clustering algorithms. Through more sophisticated clustering analysis, it not only optimizes the selection and configuration of Anchor Box, but also significantly enhances the SSD algorithm's ability to detect drone targets in complex environments, thereby further improving the detection performance. Through innovative improvements to the classic SSD algorithm, the detection efficiency of drones in low-altitude complex environments is enhanced, thereby achieving efficient and accurate recognition of drone targets; 2. In the present invention, a top-down feature information propagation mechanism is adopted, which not only retains the advantages of shallow features in locating the object position, but also ensures that the integrity and accuracy of the image semantic information can be maintained even in a complex feature fusion process by fusing deep semantic information; 3. In the present invention, in the image target detection task, multiple feature layers are processed to generate a large number of candidate frames, which are mixed with a large number of erroneous, overlapping and uncertain samples. The traditional non-maximum suppression (NMS) algorithm often causes missed detection problems due to excessive suppression when processing scenes where the distance between targets is close. To this end, the present invention introduces the DIoU-NMS algorithm, which not only measures the overlap between the candidate frame and the real frame, but also considers the distance between the center points of the two as the basis for evaluating the positional relationship. This comprehensive consideration makes the overlap evaluation between candidate frames more comprehensive and accurate, effectively reduces error suppression, and thus improves the detection accuracy and generalization ability of the SSD model in complex scenarios.

[0016] 4. In the present invention, by improving the SSD algorithm, the DBSCAN + K-Means clustering algorithm is used to optimize the AnchorBox scale, improve the matching probability between the target object and the Anchor Box, and combine the multi-branch residual unit, sub-pixel convolution, and feature pyramid to obtain multi-scale fusion features of the image. Finally, the DIoU-NMS algorithm is introduced to replace the traditional NMS algorithm to deal with the problem of redundant frames in image targets, further reduce the missed detection of targets, and thus further improve the low-altitude UAV target detection effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a diagram of the overall architecture of a model of the present invention.

[0018] Figure 2 This is a flow chart of the Anchor Box scale optimization strategy of the present invention.

[0019] Figure 3 This is a structural diagram of a multi-branch residual unit of the present invention.

[0020] Figure 4 A sub-pixel convolution operation process diagram of the present invention.

[0021] Figure 5 This is a diagram illustrating parameters of an RDIoU formula of the present invention. DETAILED DESCRIPTION

[0022] In order to more clearly illustrate the embodiments of the 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 some embodiments of the invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0023] like Figures 1 to 5 As shown in Figure 1, a low-altitude UAV target detection optimization algorithm based on deep learning includes Step 1: Train the SSD target detection model using a publicly labeled dataset to obtain the initial SSD target detection model.

[0024] In this process, a public labeled data set is selected and divided into a training set, a validation set, and a test set, wherein the training set is rotated, scaled, cropped, flipped, and other operations are performed to achieve data enhancement. The present invention uses the VGG16 convolutional neural network as the basic network of the SSD target detection model.

[0025] Step 2: Calculate the optimal aspect ratio of the Anchor Box according to the Anchor Box scale optimization strategy and generate the Anchor Box.

[0026] Among them, the present invention proposes an AnchorBox scale optimization strategy based on the fusion of DBSCAN clustering algorithm and K-means clustering algorithm. First, the DBSCAN algorithm is used to analyze the density distribution of samples in the training data set, effectively remove noise points, form new samples, and adjust the initial K value of the K-means algorithm accordingly. Subsequently, through K-means clustering, the appropriate Anchor Box scale is selected based on the key indicator of the intersection over union (IoU) between samples.

[0027] The distance formula D of the K-means algorithm in the clustering process is defined as follows: , Where: B is the sample box size; C is the cluster center generated by the K-means clustering algorithm; A is the cluster box size; IoU is the interaction ratio of the sample box size and the cluster box size, which is used as a measure of the similarity of the two box sizes.

[0028] The detailed implementation steps of the Anchor Box scale optimization strategy are as follows Figure 2 As shown: 1) Data preprocessing: First, extract the size of each image and the corresponding target box information from the training data set, which includes the coordinates of the lower left corner (Xmin, Ymin) and the upper right corner (Xmax, Ymax) of the target box. These four coordinate values ​​define the exact position of the target box in the image.

[0029] 2) Calculate the size of the target frame: Next, based on the coordinate information obtained in 1), calculate the width W and height H of each target frame. The width W is obtained by subtracting Xmax from Xmin, and the height H is obtained by subtracting Ymax from Ymin.

[0030] 3) Image and target box scaling: For unified processing, the images and their target boxes in the training dataset are adaptively scaled according to a certain ratio to obtain a new width W' and height H'. This step ensures that images and target boxes of different sizes can be analyzed at the same scale.

[0031] 4) Initialize Anchor Box: Use K-means clustering algorithm for initialization. In the scaled target box size (W', H') set, randomly select K sample points as the initial clustering centers. These K centers represent the initial K anchor boxes.

[0032] 5) Distance calculation and classification: Calculate the distance between each (W', H') in the data set and all current cluster centers, and assign each (W', H') to the nearest Anchor Box based on the minimum distance principle.

[0033] 6) Cluster center optimization: After completing the preliminary classification, the DBSCAN clustering algorithm is introduced to recalculate the cluster centers and remove noise points to optimize the cluster center position in each category.

[0034] 7) Iterative optimization: Repeat 5) and 6) until the position of the cluster center no longer changes. At this point, the obtained cluster center is determined as the final Anchor Box.

[0035] Step 3: Introduce the pyramid feature fusion module to fuse the deep features and the semantic information of the shallow feature map.

[0036] The present invention proposes a pyramid feature fusion module to achieve effective fusion of cross-level features. First, a multi-branch residual unit is introduced into the feature map to capture the multi-scale features of the image, and then the deep feature map is upsampled through sub-pixel convolution operation. Subsequently, these upsampled deep feature maps are fused with shallow feature maps to ensure that the image retains rich shallow details while also incorporating sufficient high-level semantic information.

[0037] (1) Introducing multi-branch residual units to fuse image features from different receptive fields The multi-branch residual unit replaces the traditional 3×3 convolution kernel by introducing a multi-branch convolution group. These branches are interconnected in a hierarchical structure similar to the residual network, so that the multi-scale information of the image can be more effectively utilized at a finer level. The structure of the multi-branch residual unit is shown in the figure Figure 3 shown.

[0038] The detailed process is Figure 3 It can be seen that the feature map extracted by the i-th layer is used as input (i represents conv4_3, conv7, conv8_2, conv9_2, conv10_2, conv11_2). First, after a 1×1 convolution operation, scale=4 is set, and the feature map is divided into 4 independent branch feature maps Xi (i=1,2,3,4). At this time, each Xi maintains the same width and height as the input feature map, while its number of channels is reduced to 1 / 4 of the input feature. Feature map X1 is directly output to form feature map U1, feature map X2 is output through a 3×3 convolution operation to output feature map U2, feature map X3 is first fused with the result of the convolution operation of feature map X2, and then through a 5×5 convolution operation to output feature map U3, feature map X4 is first fused with the result of the convolution operation of feature map X3, and then through a 7×7 convolution operation to output feature map U4. U1, U2, U3, and U4 are fused and then subjected to a 1×1 convolution operation to make the number of feature map channels consistent with that before entering the multi-branch residual unit, and output as the final result.

[0039] (2) Based on (1), sub-pixel convolution technology is used to upsample the feature map.

[0040] Sub-pixel convolution technology improves spatial resolution by rearranging pixel values ​​in feature map channels without introducing additional parameters or computational costs. The sub-pixel convolution operation process is shown in the figure below. Figure 4 As shown: The detailed process is Figure 4It can be seen that the feature map after the multi-branch residual unit is input as the first group of input layers (W is width, H is height, and C is the number of channels), and the feature extraction is performed through the convolution operation with a step size of 1, 3×3×r (r represents the upsampling multiple) as the second group of hidden layers, with a size of W×H×C×r. Further feature extraction is performed through the 3×3×r2 convolution operation, and the resulting feature map size is W×H×C×r×r, and the r2 channels of each pixel of the feature map are rearranged into an r×r matrix to form the third group of pixel combination layers, and finally the r×r matrices of all pixels are spliced ​​together to form a complete high-resolution image as the fourth group of output layers for output. At this time, the size of the new image is enlarged r times in width and height, with a size of W×r×H×r×C.

[0041] Step 4: Calculate the loss value through the SSD loss function.

[0042] The target loss function of the conventional SSD target detection model is the same as that of the target loss function of the present invention. The target loss function is divided into two parts: the confidence loss of the prediction box and the target category and the loss of the corresponding position regression. The specific formula is as follows: , Among them, N is the number of matches between the predicted box and the real box, x represents the matching result, c and l represent the category confidence and location information of the predicted result respectively, g represents the real box information, and the parameter α is used to adjust the ratio between the location loss and the confidence loss, which is set to 1, L conf Represents the confidence loss, specifically using Softmax Loss as the loss function, L loc Represents the position regression loss, and specifically uses SmoothL1 Loss as the loss function.

[0043] Step 5: Perform back propagation based on the loss value and update the model gradient.

[0044] By calculating the loss value obtained by the loss function, starting from the output layer, the partial derivative of each neuron is calculated reversely layer by layer according to the chain rule of the error, the adjustment amount of each weight is calculated according to the stochastic gradient descent rule, and the adjusted weights are applied to the network to update the value.

[0045] Step 6: Use the DIoU-NMS algorithm to eliminate overlapping detection frames and obtain the detection frame with the highest confidence, which is the optimal detection result.

[0046] In order to solve the problem that NMS is prone to missing targets when targets are close to each other, the present invention uses Distance-IoU (DIoU) NMS to replace the original NMS. Compared with the original NMS, the DIoU-NMS of the present invention considers the IoU value and the distance factor between the two candidate boxes, and comprehensively determines whether to retain or delete the candidate box. The formula is as follows: , Where si is the classification confidence; ε is the NMS threshold, and IoU is the intersection-over-union ratio of the sample box size and the clustering box size. is the distance between the center points of box b and box bgt; c represents the length of the diagonal of the smallest rectangle that can cover both boxes at the same time, such as Figure 5 shown.

[0047] The present invention combines the advantages of DBSCAN and K-means clustering algorithms. Through more sophisticated clustering analysis, it not only optimizes the selection and configuration of Anchor Box, but also significantly enhances the SSD algorithm's ability to detect drone targets in complex environments, thereby further improving the detection performance. Through innovative improvements to the classic SSD algorithm, the detection efficiency of drones in low-altitude complex environments is enhanced, thereby achieving efficient and accurate recognition of drone targets. The top-down feature information propagation mechanism not only retains the advantages of shallow features in locating object positions, but also ensures that the integrity and accuracy of image semantic information can be maintained even in complex feature fusion processes by fusing deep semantic information. In the image object detection task, multiple feature layers are processed to generate a large number of candidate boxes, which are mixed with a large number of erroneous, overlapping and uncertain samples. The traditional non-maximum suppression (NMS) algorithm often causes missed detection problems due to excessive suppression when processing scenes where the distance between objects is close. To this end, the present invention introduces the DIoU-NMS algorithm, which not only measures the overlap between the candidate box and the real box, but also considers the distance between the center points of the two as the basis for evaluating the positional relationship. This comprehensive consideration makes the overlap evaluation between candidate boxes more comprehensive and accurate, effectively reduces error suppression, and thus improves the detection accuracy and generalization ability of the SSD model in complex scenes.

[0048] By improving the SSD algorithm, the DBSCAN + K-Means clustering algorithm is used to optimize the Anchor Box scale, improve the matching probability between the target object and the Anchor Box, and combine the multi-branch residual unit, sub-pixel convolution, and feature pyramid to obtain multi-scale fusion features of the image. Finally, the DIoU-NMS algorithm is introduced to replace the traditional NMS algorithm to deal with the problem of redundant boxes in image targets, further reducing the missed detection of targets, thereby further improving the low-altitude UAV target detection effect.

[0049] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modification, equivalent substitution or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A low-altitude UAV target detection optimization algorithm based on deep learning, It is characterized in that The following steps are involved: Step S1: Train the SSD target detection model using a public labeled data set to obtain an initial SSD target detection model; Step S2: Calculate the optimal aspect ratio of the Anchor Box according to the Anchor Box scale optimization strategy and generate the Anchor Box; Step S3: Introduce the pyramid feature fusion module to fuse the deep features and the shallow feature map semantic information; Step S4: Calculate the loss value by using the SSD loss function; Step S5: Perform back propagation according to the loss value and update the model gradient; Step S6: Use the DIoU-NMS algorithm to eliminate overlapping detection frames and obtain the detection frame with the highest confidence, that is, the optimal detection result.

2. According to the low-altitude UAV target detection optimization algorithm based on deep learning in claim 1, It is characterized in that The model training in step S1 is as follows: Step S11: Obtain a publicly labeled data set and divide it into a training set, a validation set, and a test set; Step S12: Use the VGG16 convolutional neural network as the basic network to train the SSD target detection model to obtain the SSD target detection initial model.

3. According to the low-altitude UAV target detection optimization algorithm based on deep learning in claim 1, It is characterized in that The Anchor Box is calculated and generated in step S2 in the following way: Step S21: Data preprocessing: extract the size of each image and the corresponding target frame information from the training data set, and calculate the width and height of the target frame; Step S22: Image and target frame scaling: the image and its target frame are adaptively scaled proportionally so that the image and the target frame are analyzed at the same size; Step S23: Initialize Anchor Box: randomly select K sample points in the scaled target box size set as the initial clustering centers; Step S24: Distance calculation and classification: Calculate the distance between each target box and all cluster centers, and assign each target box to the nearest Anchor Box; Step S25: removing noise points by using DBSCAN clustering algorithm to optimize the location of cluster centers; Step S26: Repeat steps S24 to S25 until the cluster center position is stable and the final Anchor Box is generated.

4. According to the low-altitude UAV target detection optimization algorithm based on deep learning as claimed in claim 3, It is characterized in that The distance formula D of the K-means algorithm in the clustering process is: , Where B is the sample box size; C is the cluster center generated by the K-means clustering algorithm; A is the cluster box size; IoU is the interaction ratio of the sample box size and the cluster box size, which is used as a measure of the similarity of the two box sizes.

5. According to the low-altitude UAV target detection optimization algorithm based on deep learning in claim 1, It is characterized in that The step S3 is implemented in the following way: Step S31: introducing a multi-branch residual unit to fuse image features from different receptive fields; Step S32: Sub-pixel convolution improves the spatial resolution of the image by rearranging the pixel values ​​in the feature map channel through a specific convolution operation, and outputs a high-resolution image with a width and height that are r times that of the original image.

6. According to the low-altitude UAV target detection optimization algorithm based on deep learning in claim 1, It is characterized in that In step S4, the calculation is performed in the following manner: The target loss function is divided into two parts: the confidence loss of the predicted box and the target category, and the loss of the corresponding position regression.

7. According to the low-altitude UAV target detection optimization algorithm based on deep learning as claimed in claim 6, It is characterized in that The calculation formula of the objective loss function is: , Where N is the number of matches between the predicted box and the real box, x represents the matching result, c and l represent the category confidence and location information of the predicted result respectively, g represents the real box information, and the parameter α is used to adjust the ratio between the location loss and the confidence loss, which is set to 1, L conf Represents confidence loss, specifically using Softmax Loss as the loss function, L loc Represents the position regression loss, and specifically uses SmoothL1 Loss as the loss function.

8. According to the low-altitude UAV target detection optimization algorithm based on deep learning in claim 1, It is characterized in that The model gradient is updated in step S5 as follows: By calculating the loss value obtained by the loss function, starting from the output layer, the partial derivative of each neuron is calculated reversely layer by layer according to the chain rule of the error, the adjustment amount of each weight is calculated according to the stochastic gradient descent rule, and the adjusted weights are applied to the network to update the value.

9. According to the low-altitude UAV target detection optimization algorithm based on deep learning in claim 1, It is characterized in that The calculation formula of the DIoU-NMS algorithm in step S6 is: , Where si is the classification confidence; ε is the NMS threshold, and IoU is the intersection-over-union ratio of the sample box size and the clustering box size. is the distance between the center points of box b and box bgt; c represents the length of the diagonal of the smallest rectangle that can cover both boxes at the same time.

Citation Information

Patent Citations

  • Power distribution room foreign matter detection method based on SSD improvement

    CN115690584A

  • Object detection method and device and electronic equipment

    CN117576376A

  • Lightweight livestock target detection method based on dual environment perception feature pyramid

    CN118781627A