A lightweight target detection method and system for edge devices

Abstract

The application belongs to the technical field of computer vision, and proposes a lightweight target detection method and system for edge devices, which comprises the following steps: a lightweight target detection model based on a yolov5 network is constructed; the lightweight target detection model comprises a PP- LCNet backbone network, an ECA attention mechanism layer and an activation function layer; training data sets are acquired and input into the lightweight target detection model for training; a loss function is constructed, the parameters of the lightweight target detection model are updated by minimizing the loss function, and the training of the lightweight target detection model is completed; the trained lightweight target detection model is transplanted to an edge device, and a to-be-detected image is input into the trained lightweight target detection model to obtain a target detection result. The application has the characteristics of light weight and high precision, can be effectively applied to edge devices, and expands the application scenarios and practicality of edge devices.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and more specifically, to a lightweight target detection method and system for edge devices. Background Technology

[0002] With the rapid development and widespread application of computer vision technology, object detection has become a key task in many application fields, such as autonomous driving, intelligent monitoring, drone navigation, and facial recognition. However, object detection algorithms typically require a large amount of computing resources and storage space, which limits their application in resource-constrained environments.

[0003] In resource-constrained applications, such as embedded devices, mobile devices, and edge computing devices, computing resources and storage space are limited. Traditional object detection algorithms, such as deep learning-based convolutional neural network (CNN) models, have a large number of parameters and high computational complexity, making them unsuitable for direct application in these resource-constrained devices. Therefore, researching lightweight object detection algorithms has become crucial for solving this problem. Current lightweight object detection algorithms rely on pruning, quantization, and distillation, but these methods significantly reduce model accuracy, resulting in poor performance in practical object detection applications. Summary of the Invention

[0004] To overcome the shortcomings of existing lightweight target detection algorithms, such as high computational resource consumption and low accuracy, this invention provides a lightweight target detection method and system for edge devices.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0006] A lightweight target detection method for edge devices includes the following steps:

[0007] A lightweight object detection model based on the YOLOv5 network is constructed. The lightweight object detection model includes a PP-LCNet backbone network, an ECA (Efficient Channel Attention) attention mechanism layer, and an activation function layer.

[0008] Obtain the training dataset and input it into the lightweight object detection model for training;

[0009] Construct a loss function, and update the parameters of the lightweight object detection model by minimizing the loss function to complete the training of the lightweight object detection model;

[0010] The trained lightweight object detection model is ported to an edge device, and the image to be detected is input into the trained lightweight object detection model to obtain the object detection result.

[0011] As a preferred embodiment, the PP-LCNet backbone network includes at least 5 stacked depthwise separable convolutional layers; wherein, the output of the second depthwise separable convolutional layer is connected to a first CBH module, and the third depthwise separable convolutional layer is connected to a second CBH module;

[0012] The output of the last depthwise separable convolutional layer is connected to the input of the ECA attention mechanism layer, and the output of the ECA attention mechanism layer is connected to the CSP2_1 module; the output of the CSP2_1 module is connected to the third CBH module;

[0013] The activation function layer is connected to the ends of the first CBH module, the second CBH module, the third CBH module, and the CSP2_1 module;

[0014] The first CBH module, the second CBH module, and the third CBH module are concatenated with each other and passed through convolutional structures to output the first feature, the second feature, and the third feature, respectively, which are used to generate target detection results through a classifier.

[0015] As a preferred embodiment, the first CBH module, the second CBH module, and the third CBH module include a convolutional layer, a BN layer, and the activation function layer connected in sequence.

[0016] As a preferred embodiment, in the CSP2_1 module, its input passes through two CBH modules and a convolutional layer connected in sequence, and is then concatenated with the output of a single convolutional layer, and then passes through a BN layer and the activation function layer in sequence.

[0017] As a preferred embodiment, the activation function layer includes the H-Swish activation function; its expression is:

[0018]

[0019] in, Indicates input x After activation by the ReLU function and truncation, i.e. .

[0020] As a preferred embodiment, the loss function includes the EIoU loss function; its expression is:

[0021]

[0022] in, This represents the IoU (Intersection over Union) loss. Indicates distance loss. Indicates aspect ratio loss; These are the weighting coefficients; The predicted value for the center point of the bounding box. This represents the actual value of the center point of the bounding box. The scaling factor is the center point scaling factor. To predict the box width, This is the actual frame width; This is the width scaling factor; To predict the bounding box height, This is the actual bounding box height. This is the height scaling factor.

[0023] As a preferred embodiment, the step of porting the trained lightweight target detection model to an edge device includes:

[0024] The trained lightweight object detection model is extracted into ONNX format;

[0025] The lightweight target detection model in ONNX format was optimized using TensorRT and converted into an engine file.

[0026] The engine file is loaded on the edge device side.

[0027] Furthermore, the present invention also proposes a lightweight target detection system for edge devices, comprising:

[0028] The image acquisition module is used to acquire the training image set and the image to be detected;

[0029] Memory on which computer-readable instructions are stored;

[0030] A processor for executing the computer-readable instructions; when the computer-readable instructions are executed by the processor, the processor causes the processor to perform the steps of the lightweight target detection method proposed in this invention;

[0031] The display module is used to display the target detection results.

[0032] Furthermore, the present invention also proposes a computer device including a memory and a processor, wherein the memory stores computer-readable instructions, wherein when the computer-readable instructions are executed by the processor, the processor causes the processor to perform the steps of the lightweight target detection method proposed in the present invention.

[0033] Furthermore, the present invention also proposes a storage medium storing computer-readable instructions thereon, wherein the computer-readable instructions, when executed by a processor, implement the steps of the lightweight target detection method proposed in the present invention.

[0034] Compared with existing technologies, the beneficial effects of the technical solution of this invention are as follows: This invention replaces the backbone network of the initial YOLOv5 network architecture with the lightweight network PP-LCNet, constructing a lightweight object detection model based on the YOLOv5 network. While ensuring the accuracy of the initial YOLOv5 network architecture, it significantly improves the inference speed of the model and reduces the model's parameter size and computational load. This invention also incorporates the ECA attention mechanism, which has good generalization ability in object detection and instance segmentation tasks, and further enhances the feature representation ability of the backbone network with lower computational complexity. Attached Figure Description

[0035] Figure 1 This is a flowchart of the lightweight target detection method in Example 1.

[0036] Figure 2 This is an architecture diagram of the lightweight target detection model in Example 1.

[0037] Figure 3 This is an architecture diagram of the lightweight target detection system in Example 2. Detailed Implementation

[0038] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent; it is understandable that certain well-known structures and their descriptions may be omitted in the drawings for those skilled in the art.

[0039] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0040] Example 1

[0041] This embodiment proposes a lightweight target detection method for edge devices, such as... Figure 1 The diagram shown is a flowchart of the lightweight target detection method in this embodiment.

[0042] The lightweight target detection method proposed in this embodiment includes the following steps:

[0043] S1. Construct a lightweight object detection model based on the YOLOv5 network.

[0044] The lightweight object detection model in this embodiment includes the PP-LCNet backbone network, the ECA attention mechanism layer, and the activation function layer.

[0045] S2. Obtain the training dataset and input it into the lightweight object detection model for training.

[0046] S3. Construct a loss function, and update the parameters of the lightweight object detection model by minimizing the loss function to complete the training of the lightweight object detection model.

[0047] S4. The trained lightweight target detection model is ported to the edge device, and the image to be detected is input into the trained lightweight target detection model to obtain the target detection result.

[0048] In this embodiment, the backbone network of the initial YOLOv5 network architecture is replaced with the lightweight network PP-LCNet. This lightweight network is implemented based on PaddlePaddle and can improve the performance of the model on multiple tasks. It can also improve the model accuracy while keeping the latency almost unchanged. While ensuring the accuracy of the initial YOLOv5 network architecture, it can significantly improve the inference speed of the model and reduce the model's parameter size and computational load.

[0049] Furthermore, this embodiment incorporates an ECA attention mechanism layer into the lightweight object detection model based on the YOLOv5 network. ECA attention is an attention mechanism used in computer vision tasks, aiming to improve the expressive and perceptual capabilities of convolutional neural networks (CNNs) while reducing computational complexity. ECA attention achieves this through adaptive local interactions along the channel dimension. Specifically, it models the inter-channel dependencies by introducing a one-dimensional convolution operation. This convolution operation can be viewed as an adaptive sliding window averaging, averaging the responses of adjacent channels at each location and then using the result to adjust the weights of each channel.

[0050] In this embodiment, the addition of the ECA attention mechanism layer only increases a small number of parameters, but can achieve significant performance gains. It can effectively avoid dimensionality reduction, capture cross-channel interaction information, and aims to ensure information efficiency and effectiveness. It can improve the performance of various deep CNN architectures and has good generalization ability in object detection and instance segmentation tasks. It can further enhance the feature representation ability of the backbone network with lower computational complexity.

[0051] The lightweight target detection method in this embodiment can be effectively applied to edge devices, expanding the application scenarios and practicality of edge devices.

[0052] In an alternative embodiment, the ECA attention mechanism layer generates channel attention through fast one-dimensional convolution, and its kernel size is adaptively determined through a nonlinear mapping of the channel dimension.

[0053] In one optional embodiment, the PP-LCNet backbone network includes at least 5 stacked depthwise separable convolutional layers; wherein the output of the second depthwise separable convolutional layer is connected to a first CBH module, and the third depthwise separable convolutional layer is connected to a second CBH module.

[0054] The output of the last depthwise separable convolutional layer is connected to the input of the ECA attention mechanism layer, and the output of the ECA attention mechanism layer is connected to the CSP2_1 module; the output of the CSP2_1 module is connected to the third CBH module.

[0055] The activation function layer is connected to the ends of the first CBH module, the second CBH module, the third CBH module, and the CSP2_1 module.

[0056] The first CBH module, the second CBH module, and the third CBH module are concatenated with each other and passed through convolutional structures to output the first feature, the second feature, and the third feature, respectively, which are used to generate target detection results through a classifier.

[0057] like Figure 2 The diagram shown is an architecture diagram of the lightweight target detection model in this embodiment.

[0058] The lightweight object detection model in this embodiment is a YOLOv5 + PP-LCNet structure, where the backbone network of the initial YOLOv5 network architecture is replaced with the lightweight PP-LCNet network. PP-LCNet uses depthwise separable convolution (DepthSepConv) as its basic module, and builds a backbone network similar to MobileNetV1 by stacking DepthSepConv modules. Then, the backbone networks are combined to construct PP-LCNet.

[0059] PP-LCNet is a deep neural network for perceptual image quality assessment. It predicts the perceptual quality of an image by combining perceptual features and contextual information. The network structure employs a branching and fusion design, utilizing residual blocks and attention mechanisms to extract feature representations of the image.

[0060] Furthermore, in an optional embodiment, the first CBH module, the second CBH module, and the third CBH module include a convolutional layer, a BN layer, and the activation function layer connected in sequence.

[0061] Furthermore, in an optional embodiment, in the CSP2_1 module, its input passes through two CBH modules and a convolutional layer connected in sequence, and is then concatenated with its output after passing through a single convolutional layer, and then passes through a BN layer and the activation function layer in sequence.

[0062] As an example, the PP-LCNet backbone network of this embodiment includes five stacked DepthSepConv layers. The output of the second DepthSepConv layer is connected to the CBH-1 module, the output of the third DepthSepConv layer is connected to the CBH-2 module, the output of the fifth DepthSepConv layer is connected to the ECA attention mechanism layer, the output of the ECA attention mechanism layer is connected to the CSP2_1 module, and the output of the CSP2_1 module is connected to the CBH-3 module.

[0063] The output of CBH-2 is concatenated with the output of CBH-3 after upsampling, then passes through the CSP2_1 module, CBH-4 and the upsampling layer in sequence, and is concatenated with the output of CBH-1. Finally, it passes through the CSP2_1 module, CBH-5 and the convolutional layer in sequence to output the third feature.

[0064] The outputs of CBH-4 and CBH-5 are concatenated and then passed through the CSP2_1 module, CBH-6, and convolutional layers in sequence to output the second feature.

[0065] The outputs of CBH-3 and CBH-6 are concatenated, and then passed through the CSP2_1 module, CBH-6 and convolutional layers in sequence to output the first feature.

[0066] In this embodiment, CBH-1, CBH-2, CBH-3, CBH-4, CBH-5, and CBH-6 all have the same structure, including a convolutional layer, a BN layer, and an activation function layer connected in sequence.

[0067] Furthermore, in an optional embodiment, the activation function layer includes the H-Swish activation function. Its expression is:

[0068]

[0069] in, Indicates input x After activation by the ReLU function and truncation, i.e. .

[0070] Swish is an activation function with smooth nonlinear characteristics and exhibits better performance than other commonly used activation functions in certain situations. The H-Swish (Hard Swish) function used in this embodiment is an improved activation function based on Swish, further improving computational efficiency by reducing the multiplication operations in the Swish function. The H-Swish function, similar to the Swish function, has nonlinear characteristics, but it reduces computational complexity by replacing the sigmoid and multiplication operations in the Swish function with the ReLU6 function.

[0071] In this embodiment, the H-Swish activation function is used with ReLU6 for truncation and scaling, which significantly reduces computational overhead and improves the speed of forward propagation. Meanwhile, the H-Swish function still retains non-linear characteristics, providing a non-linear response to the input, which helps the model to have stronger expressive power when learning complex features. Furthermore, the H-Swish function has adaptive properties similar to the Swish function; when the input is negative, it can compress it while retaining some negative information.

[0072] Furthermore, in an optional embodiment, the loss function includes the EIoU loss function. Its expression is:

[0073]

[0074] in, Indicates IoU loss. Indicates distance loss. Indicates aspect ratio loss; These are the weighting coefficients; The predicted value for the center point of the bounding box. This represents the actual value of the center point of the bounding box. The scaling factor is the center point scaling factor. To predict the box width, This is the actual frame width; This is the width scaling factor; To predict the bounding box height, This is the actual bounding box height. This is the height scaling factor.

[0075] In this embodiment, the initial loss function CIOU of the YOLOv5 network model is replaced with the EIOU loss function, which includes overlap loss, center distance loss, and width-height loss. The first two parts continue the methods in the CIOU loss function, but the width-height loss directly minimizes the difference in width and height between the target box and the anchor box, resulting in faster convergence.

[0076] While CIoU loss considers factors such as overlapping area, center distance, and aspect ratio in bounding box regression, it hinders the model's effective optimization of similarity by reflecting differences in aspect ratio in its formula, rather than the true differences between width and height and their confidence scores. To address this issue, EIoU loss separates aspect ratio from CIoU and incorporates Focal focusing on high-quality anchor boxes.

[0077] In this embodiment, the EIoU loss function corrects for the shortcomings of the CIoU metric by introducing an additional offset. Specifically, the EIoU loss function models the target localization problem as an estimation problem of the optimal offset. By minimizing the error between the corrected IoU and the true bounding box, more accurate target localization results can be obtained, which helps to improve the performance of the target detection task.

[0078] As an example, the calculation process of the EIoU loss function includes the following steps:

[0079] (1) Calculate the traditional IoU, which is the area of ​​the intersection between the predicted bounding box and the true bounding box divided by the area of ​​their union;

[0080] (2) Calculate the center offset of the bounding box, which is used to measure the error of the center position of the predicted bounding box relative to the true bounding box;

[0081] (3) Calculate the width offset and height offset of the bounding box to measure the scale error of the predicted bounding box's width and height relative to the true bounding box;

[0082] (4) Based on the center offset, width offset and height offset, the EIoU loss value is obtained.

[0083] Furthermore, as an example, the performance of the YOLOv5 network and the lightweight object detection model (LC-YoLOv5) of this embodiment is compared using the Coco-Nano dataset. Specifically, a script was written to select ten categories for indoor scenes from the Coco-Nano dataset and remove labels from other categories to obtain the Coco-Nano dataset, enabling object detection of household items in indoor scenes. The ten categories in the Coco-Nano dataset are: bench, suitcase, bottle, cub, couch, bed, dining table, laptop, microwave, and oven, with a total data size of 24912. The test results of LC-YoLOv5 and YOLOv5s on the Coco-Nano dataset at the same scale are shown in Tables 1 and 2.

[0084] Table 1. Model Performance Comparison Results

[0085]

[0086] Table 2 Comparison of Reasoning Speed

[0087]

[0088] As shown in the table above, the lightweight target detection model based on the YOLOv5 network in this embodiment has a smaller model size, consumes less computational resources, and can still guarantee a certain level of detection accuracy and inference speed compared to the traditional YOLOv5 network.

[0089] Furthermore, in an optional embodiment, porting the trained lightweight object detection model to an edge device includes the following steps:

[0090] S4.1 Extract the trained lightweight object detection model into ONNX format;

[0091] S4.2 Optimize the lightweight target detection model in ONNX format using TensorRT and convert it into an engine file;

[0092] S4.3 Load the engine file on the edge device side, and input the image to be detected into the trained lightweight target detection model to obtain the target detection result.

[0093] ONNX defines a set of environment- and platform-independent standard formats to enhance model interactivity. In this embodiment, the trained lightweight object detection model is extracted into the ONNX format for easier porting in the next step.

[0094] TensorRT is a deep learning optimizer and inference engine launched by NVIDIA for high-performance inference. It is used to optimize deep learning models for specific hardware platforms (such as NVIDIA GPUs) and provide faster and more efficient inference performance.

[0095] As an example, the steps for optimizing the lightweight object detection model in ONNX format using TensorRT include:

[0096] (1) Compile ONNX-TensorRT: Import model data, define and optimize the network of the model through TensorRT, analyze the structure of the model, identify and trim unnecessary computations, merge and fuse network layers to reduce memory usage and computation, and improve inference performance.

[0097] (2) ONNX to engine: TensorRT builds the inference engine and serializes and saves the inference engine to the hard disk. The inference engine is a highly optimized computation graph optimized for a specific hardware platform.

[0098] (3) TensorRT acceleration: TensorRT provides optimization and inference acceleration for deep learning models. It reduces computation and memory overhead through techniques such as pruning, fusion, and quantization, and provides fast inference performance on specific hardware using a highly optimized inference engine.

[0099] Furthermore, the engine file is loaded into the code on the edge device side, compiled into an executable file, and target detection in indoor scenes is implemented on the edge device side.

[0100] Example 2

[0101] This embodiment proposes a lightweight target detection system for edge devices, applying the lightweight target detection method proposed in Embodiment 1. For example... Figure 3 The diagram shown is an architecture diagram of the lightweight target detection system in this embodiment.

[0102] The lightweight target detection system for edge devices proposed in this embodiment includes:

[0103] The image acquisition module is used to acquire the training image set and the image to be detected;

[0104] Memory on which computer-readable instructions are stored;

[0105] A processor for executing the computer-readable instructions; when the computer-readable instructions are executed by the processor, the processor causes the processor to perform the steps of the lightweight target detection method proposed in Embodiment 1;

[0106] The display module is used to display the target detection results.

[0107] As an example, the Jetson Nano Developer Kit B01 was selected as the processor, a 64GB TF memory card as the storage, a MIPI-CSI high-definition camera as the image stimulation module, and a display as the display module.

[0108] In the specific implementation process, the system image corresponding to the development board version from the NVIDIA official website was downloaded and burned to a TF memory card. A monitor was connected to the development board, the TF memory card was inserted, and the system was installed, along with the necessary environment for the code, such as PyTorch and TorchVision. OnNX-TensorRT was compiled on the development board to prepare for TensorRT acceleration. The trained lightweight object detection model data, converted to ONNX format, was converted into an engine file, adapted for TensorRT acceleration, and further compiled to generate an executable file. Finally, a MIPI-CSI high-definition camera was installed on the development board to perform object detection in an indoor environment to test the accuracy.

[0109] It is understood that the system in this embodiment corresponds to the method in Embodiment 1 above, and the options in Embodiment 1 above are also applicable to this embodiment, so they will not be described again here.

[0110] Example 3

[0111] This embodiment proposes a computer device, including a memory and a processor. The memory stores computer-readable instructions, wherein when the computer-readable instructions are executed by the processor, the processor performs the steps of the lightweight target detection method proposed in Embodiment 1.

[0112] Example 4

[0113] This embodiment proposes a storage medium storing computer-readable instructions, wherein when the computer-readable instructions are executed by a processor, they implement the steps of the lightweight target detection method proposed in Embodiment 1.

[0114] By way of example, the storage medium includes, but is not limited to, USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks or optical disks, and other media capable of storing program code.

[0115] By way of example, the instructions, programs, code sets, or instruction sets may be implemented using conventional programming languages.

[0116] By way of example, the processor includes, but is not limited to, smartphones, personal computers, servers, network devices, etc., for performing all or part of the steps of the lightweight target detection method described in Example 1.

[0117] The terminology used in the accompanying drawings is for illustrative purposes only and should not be construed as limiting the scope of this patent.

[0118] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A lightweight target detection method for edge devices, characterized in that, Includes the following steps: A lightweight object detection model based on the YOLOv5 network is constructed; the lightweight object detection model includes a PP-LCNet backbone network, an ECA attention mechanism layer, and an activation function layer; Obtain the training dataset and input it into the lightweight object detection model for training; Construct a loss function, and update the parameters of the lightweight object detection model by minimizing the loss function to complete the training of the lightweight object detection model; The trained lightweight object detection model is ported to an edge device, and the image to be detected is input into the trained lightweight object detection model to obtain the object detection result. The PP-LCNet backbone network includes at least 5 stacked depthwise separable convolutional layers; wherein the output of the second depthwise separable convolutional layer is connected to a first CBH module, and the third depthwise separable convolutional layer is connected to a second CBH module. The output of the last depthwise separable convolutional layer is connected to the input of the ECA attention mechanism layer, and the output of the ECA attention mechanism layer is connected to the CSP2_1 module; the output of the CSP2_1 module is connected to the third CBH module; The activation function layer is connected to the ends of the first CBH module, the second CBH module, the third CBH module, and the CSP2_1 module; The first CBH module, the second CBH module, and the third CBH module are concatenated with each other and passed through convolutional structures, respectively outputting the first feature, the second feature, and the third feature, which are used to generate target detection results through a classifier; The first CBH module, the second CBH module, and the third CBH module include a convolutional layer, a BN layer, and the activation function layer connected in sequence; in the CSP2_1 module, its input passes through two CBH modules and a convolutional layer connected in sequence, is concatenated with the output of a single convolutional layer, and then passes through a BN layer and the activation function layer in sequence.

2. The lightweight target detection method according to any one of claims 1, characterized in that, The activation function layer includes the H-Swish activation function; its expression is: H_Swish(x) = x * relu6(x + 3) / 6 Here, relu6(x) represents the input x being processed by the ReLU activation function and then truncated, i.e., relu6(x) = min(max(0,x),6).

3. The lightweight target detection method according to claim 2, characterized in that, The loss function includes the EIoU loss function; its expression is: Among them, L IOU Indicates IoU loss, L dis L represents the distance loss. asp This represents the aspect ratio loss; ρ is the weighting coefficient; b is the predicted value of the bounding box center point. gt Here, c represents the true value of the bounding box center point, c is the center point scaling factor, and w is the predicted box width. gt c is the actual frame width; w h is the width scaling factor; h is the height of the prediction box. gt c is the actual bounding box height. h This is the height scaling factor.

4. The lightweight target detection method according to claim 3, characterized in that, The process of porting the trained lightweight object detection model to an edge device includes: The trained lightweight object detection model is extracted into ONNX format; The lightweight target detection model in ONNX format was optimized using TensorRT and converted into an engine file. The engine file is loaded on the edge device side.

5. A lightweight target detection system for edge devices, characterized in that, include: The image acquisition module is used to acquire the training image set and the image to be detected; Memory on which computer-readable instructions are stored; A processor for executing the computer-readable instructions; when the computer-readable instructions are executed by the processor, the processor causes the processor to perform the steps of the lightweight target detection method as described in any one of claims 1 to 4; The display module is used to display the target detection results.

6. A computer device comprising a memory and a processor, wherein the memory stores computer-readable instructions, characterized in that, When the computer-readable instructions are executed by the processor, the processor performs the steps of the lightweight target detection method as described in any one of claims 1 to 4.

7. A storage medium having computer-readable instructions stored thereon, characterized in that, When the computer-readable instructions are executed by a processor, they implement the steps of the lightweight target detection method as described in any one of claims 1 to 4.

Images