A target detection method and system based on LCSD-YOLO network

Abstract

The application discloses a target detection method and system based on an LCSD-YOLO network, and the method comprises the following steps: acquiring target image data and performing data preprocessing to obtain a preprocessed image data set; introducing a PP-LCNet network architecture, a SDDS module and an attention-based internal scale feature interaction module based on a YOLO11 network model to construct an LCSD-YOLO network model; and performing target recognition detection on the preprocessed image data set based on the LCSD-YOLO network model to obtain a target recognition detection result. The application can improve the small target detection capability of the model, the extraction precision and efficiency of target features, and the precision and efficiency of small target detection. The application can be widely applied to the technical field of target recognition detection as a target detection method and system based on an LCSD-YOLO network.

Description

Technical Field

[0001] The present invention relates to the technical field of target recognition and detection, and in particular to a target detection method and system based on an LCSD-YOLO network. Background Art

[0002] At present, deep learning-based methods have significantly optimized detection performance by improving hardware computing power. However, their inherent defects such as large number of model parameters and high computational complexity lead to the bottleneck of insufficient real-time performance when deployed on edge devices. However, while existing lightweight technologies can achieve model size compression through structural optimization methods such as parameter pruning, low-rank decomposition, compact convolution design, and knowledge distillation, they suffer from the drawback of impaired feature expression capabilities, resulting in reduced small target recognition accuracy. Furthermore, the compression rate is not linearly correlated with improved computational efficiency. Traditional compression cannot meet the requirements of high precision, real-time performance, and low latency in practical applications. For small target detection, such as intelligent identification of whether workers working at heights are wearing helmets, existing YOLO series algorithms (such as YOLO11n) perform well in terms of speed and accuracy. However, images of workers working at heights captured by ground-based cameras with upward-looking images often have a low target pixel ratio (typically less than 0.1% of the image area). To address this, recent research has gradually introduced enhanced networks and feature fusion techniques to further improve detection robustness. Various attention mechanisms have also been added to improve the accuracy of small target detection. However, this introduces a large number of parameters into the model, resulting in a decrease in detection speed and impacting its real-time performance in practical applications. Summary of the Invention

[0003] In order to solve the above technical problems, the purpose of the present invention is to provide a target detection method and system based on the LCSD-YOLO network, which can improve the small target detection ability of the model and the extraction accuracy and efficiency of target features, and improve the accuracy and efficiency of small target detection.

[0004] The first technical solution adopted by the present invention is: a target detection method based on the LCSD-YOLO network, comprising the following steps:

[0005] Acquire target image data and perform data preprocessing to obtain a preprocessed image data set;

[0006] Based on the YOLO11 network model, the PP-LCNet network architecture, SDDS module and attention-based internal scale feature interaction module are introduced to build the LCSD-YOLO network model;

[0007] Based on the LCSD-YOLO network model, target recognition detection is performed on the preprocessed image dataset to obtain the target recognition detection results.

[0008] Further, the step of obtaining target image data and performing data preprocessing to obtain a preprocessed image data set specifically includes:

[0009] obtaining target image data sets by image acquisition of the target object through the camera;

[0010] naming the target image data set according to the format of the Pascal VOC data set, creating a target image data file with three folders of Annotations, ImageSets and JPEGImages;

[0011] performing data labeling processing on the target image data file to obtain a preprocessed image data set.

[0012] Further, the LCSD-YOLO network model specifically includes a backbone network module, a neck network and a head network, the backbone network module, the neck network and the head network are connected in sequence, wherein:

[0013] the backbone network module adopts a PP-LCNet network architecture, the backbone network module includes a first convolution module, an SDDS module, a depth separable convolution module and an attention-based internal scale feature interaction module, the SDDS module includes a first SDDS module, a second SDDS module and a third SDDS module, and the depth separable convolution module includes a first depth separable convolution module, a second depth separable convolution module and a third depth separable convolution module;

[0014] the neck network includes a DySample module, a Concat module, a C3K2 module, an SE module and a convolution module, the DySample module includes a first DySample module and a second DySample module, the Concat module includes a first Concat module, a second Concat module, a third Concat module and a fourth Concat module, the first C3K2 module, the second C3K2 module, the third C3K2 module and the fourth C3K2 module, and the convolution module includes a second convolution module and a third convolution module;

[0015] the head network includes a first detection head, a second detection head and a third detection head.

[0016] Further, the LCSD-YOLO network model constructs a target recognition loss function by introducing a non-monotonic focusing mechanism, and the expression of the target recognition loss function is specifically as follows:

[0017] L WIoU =r×L WIoUv1

[0018]

[0019] In the above formula, L WIoU represents the target recognition loss function, r represents the non-monotonic focusing coefficient, L WIoUv1 represents the index function of the degree of overlap between the predicted frame and the real frame, β represents the application of WIoUv1 with dynamic non-monotonic FM to WIoUv3, α and δ represent variable hyperparameters, represents the monotonic focusing coefficient, L IoU represents the intersection ratio loss function of the predicted frame and the real frame, and IoU represents the intersection ratio of the predicted frame and the real frame.

[0020] Further, the step of performing target recognition detection on the preprocessed image data set based on the LCSD-YOLO network model to obtain the target recognition detection result comprises:

[0021] inputting the preprocessed image data set into the LCSD-YOLO network model;

[0022] performing feature recognition processing on the preprocessed image data set based on the backbone network module of the LCSD-YOLO network model to obtain target image features;

[0023] performing multi-scale feature fusion processing on the target image features based on the neck network of the LCSD-YOLO network model to obtain fused target image features;

[0024] performing decoding prediction on the fused target image features based on the head network of the LCSD-YOLO network model to obtain the target recognition detection result.

[0025] Further, the step of performing feature recognition processing on the preprocessed image data set based on the backbone network module of the LCSD-YOLO network model to obtain target image features comprises:

[0026] inputting the preprocessed image data set into the backbone network module of the LCSD-YOLO network model;

[0027] performing channel adjustment and feature extraction on the preprocessed image data set based on the first convolution module of the backbone network module to obtain a preliminary target feature image;

[0028] performing convolution and spatial channel dimension conversion processing on the preliminary target feature image based on the SDDS module of the backbone network module to obtain a first intermediate target feature image;

[0029] performing depth separable convolution processing on the intermediate target feature image based on the depth separable convolution module of the backbone network module to obtain a second intermediate target feature image;

[0030] The attention-based internal scale feature interaction module based on the backbone network module performs multi-scale feature interaction redundant processing on the second intermediate target feature image to obtain the target image features.

[0031] Furthermore, the SDDS module based on the backbone network module performs convolution and spatial channel dimension conversion on the preliminary target feature image to obtain a first intermediate target feature image, which specifically includes:

[0032] Input the preliminary target feature image into the SDDS module of the backbone network module, wherein the SDDS module includes a depth-wise convolution module, a point-wise convolution module, and a spatial depth conversion module;

[0033] Based on the depth-wise convolution module of the SDDS module, the preliminary target feature image is subjected to depth-wise convolution processing to obtain the target feature image after the first convolution;

[0034] Based on the point-by-point convolution module of the SDDS module, channel aggregation processing is performed on the target feature image after the first convolution to obtain the target feature image after the second convolution;

[0035] Based on the spatial depth conversion module of the SDDS module, the target feature image after the second convolution is connected along the channel dimension to obtain the first intermediate target feature image.

[0036] Furthermore, the neck network based on the LCSD-YOLO network model performs multi-scale feature fusion processing on the target image features to obtain the fused target image features, which specifically includes:

[0037] Input the target image features into the neck network of the LCSD-YOLO network model;

[0038] The DySample module based on the neck network performs feature reconstruction on the target image features to obtain the reconstructed target image features;

[0039] The Concat module based on the neck network performs feature splicing on the reconstructed target image features to obtain the spliced ​​target image features;

[0040] The C3K2 module based on the neck network performs multi-level spatial feature extraction on the spliced ​​target image features to obtain deeper target image features;

[0041] Based on the SE module of the neck network, the channel feature weight adjustment processing is performed on the deeper target image features to obtain the adjusted target image features;

[0042] The convolution module based on the neck network performs convolution processing on the adjusted target image features and outputs the fused target image features.

[0043] The second technical solution adopted by the present invention is: a target detection system based on the LCSD-YOLO network, comprising:

[0044] The first module is used to obtain target image data and perform data preprocessing to obtain a preprocessed image data set;

[0045] The second module is used to build the LCSD-YOLO network model based on the YOLO11 network model by introducing the PP-LCNet network architecture, the SDDS module, and the attention-based internal scale feature interaction module;

[0046] The third module is used to perform target recognition detection on the preprocessed image dataset based on the LCSD-YOLO network model to obtain the target recognition detection results.

[0047] The beneficial effects of the method and system of the present invention are as follows: the present invention acquires target image data and performs data preprocessing, and then introduces the PP-LCNet network architecture, the SDDS module and the attention-based internal scale feature interaction module based on the YOLO11 network model to construct the LCSD-YOLO network model, and adopts the lightweight backbone network PP-LCNet to replace the original yolo11n backbone network, which can reduce the model complexity and reduce the computing resource requirements. Furthermore, by integrating the deep separable downsampling convolution SDDS with spatial depth conversion, the loss of fine-grained information is reduced and the learning efficiency of feature representation is improved, which significantly improves the problem of insufficient small target detection capability caused by the impaired feature expression capability of the direct compression model. Further, by performing intra-scale interaction through the attention-based internal scale feature interaction module, the computational cost is further reduced while the model feature extraction capability is improved, and the problem of insufficient small target detection accuracy caused by the damage to the feature pyramid of the lightweight model is effectively solved. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is a flowchart of the steps of a target detection method based on the LCSD-YOLO network of the present invention;

[0049] Figure 2 It is a structural block diagram of a target detection system based on LCSD-YOLO network of the present invention;

[0050] Figure 3 Schematic diagram of the structure of the LCSD-YOLO network model provided by a specific embodiment of the present invention;

[0051] Figure 41 is a parameter diagram of the PP-LCNet network architecture provided by a specific embodiment of the present invention;

[0052] Figure 5 1 is a schematic diagram of an example of depth-wise convolution provided by a specific embodiment of the present invention;

[0053] Figure 6 is a schematic diagram of an example of point-by-point convolution provided by a specific embodiment of the present invention;

[0054] Figure 7 is a schematic diagram of an example of spatial depth conversion provided by a specific embodiment of the present invention;

[0055] Figure 8 is a schematic diagram of an example of a DySample module provided in a specific embodiment of the present invention;

[0056] Figure 9 It is a schematic diagram of the target recognition detection result provided by a specific embodiment of the present invention. DETAILED DESCRIPTION

[0057] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The step numbers in the following embodiments are provided for ease of description only and do not limit the order of the steps. The order of execution of the steps in the embodiments can be adaptively adjusted based on the understanding of those skilled in the art.

[0058] First, it should be noted that the embodiment of the present invention is directed to target detection at high altitudes. The following specific implementation method of the embodiment of the present invention is described using the intelligent recognition of whether a worker working at high altitudes is wearing a safety helmet.

[0059] Reference Figure 1 , the present invention provides a target detection method based on LCSD-YOLO network, the method comprising the following steps:

[0060] S100, acquiring target image data and performing data preprocessing to obtain a preprocessed image data set;

[0061] Specifically, the target object is imaged through a camera to obtain a target image dataset; the target image dataset is named according to the format of the PascalVOC dataset, and a target image data file with three folders: Annotations, ImageSets, and JPEGImages is created; the target image data file is subjected to data labeling processing to obtain a preprocessed image dataset.

[0062] In this example, a self-made dataset was used. The experimenters used an industrial camera to capture images of the target objects and named the captured images according to the format of the Pascal VOC dataset. At the same time, three folders named Annotations, ImageSets, and JPEGImages were created. The image annotation tool LabelImg was used to mark the targets in the captured images, and the category labels were humans and helmets, respectively.

[0063] Because this dataset cannot meet the requirement of 2000 images required to identify two categories of targets, the Augmentor image data augmentation library is used to amplify the images. The experimenters select the image save path and the tag information XML file path, and formulate the output path of the amplified image and XML file. They select the required image enhancers such as brightness, cropping, Gaussian noise, etc., and select the number of amplifications and the amplification method (sequential, combined, random, etc.) to amplify the image to meet the required recognition requirements.

[0064] Since the embodiment of the present invention uses the YOLO series neural network, it is necessary to convert the amplified images and label files into the YOLO format and then divide the data set into a training set and a validation set of 75% and 25% respectively.

[0065] After building the experimental environment and preprocessing the dataset, we next created the corresponding dataset configuration file. The dataset configuration file contains the paths to the training and validation sets, as well as the dataset category information. We then modified the YOLO11 default.yaml file based on the experimental requirements, adjusting the input image size to 640×640. This model was trained for 400 rounds, with a batch size of 64 and two object categories. The Mosaic data augmentation method was used with a probability of 1, and the copy-paste data augmentation method was used with a probability of 0.5. The following environment information was used: cuda 11.8, deep learning framework pytorch 2.3.0, Intel Core i9-13900KS CPU, 128GB of RAM, and an NVIDIA GeForce RTX 4090 GPU with 24GB of video memory.

[0066] S200, based on the YOLO11 network model, introduces the PP-LCNet network architecture, SDDS module and attention-based internal scale feature interaction module to build the LCSD-YOLO network model;

[0067] Specifically, if Figure 3 As shown, the LCSD-YOLO network model specifically includes a backbone network module, a neck network and a head network, and the backbone network module, the neck network and the head network are connected in sequence, wherein:

[0068] The backbone network module adopts the PP-LCNet network architecture, and the backbone network module includes a first convolution module, an SDDS module, a depth-separable convolution module and an attention-based internal scale feature interaction module. The SDDS module includes a first SDDS module, a second SDDS module and a third SDDS module. The depth-separable convolution module includes a first depth-separable convolution module, a second depth-separable convolution module and a third depth-separable convolution module.

[0069] The neck network includes a DySample module, a Concat module, a C3K2 module, an SE module and a convolution module, the DySample module includes a first DySample module and a second DySample module, the Concat module includes a first Concat module, a second Concat module, a third Concat module and a fourth Concat module, the first C3K2 module, the second C3K2 module, the third C3K2 module and the fourth C3K2 module, and the convolution module includes a second convolution module and a third convolution module;

[0070] The head network includes a first detection head, a second detection head and a third detection head.

[0071] In this embodiment, the LCSD-YOLO network model is based on the YOLO11 network, and the overall structure consists of three parts: the backbone network (Backbone), the neck network (Neck) and the head network (Head). First, by replacing the backbone network with the lightweight network PP-LCNet, the complexity and parameter amount of the model are significantly reduced with little loss of accuracy. To further optimize the performance, the SDDS module is introduced in PP-LCNet to replace the original downsampling convolution, and the DepthSepConv of the last layer is replaced by the AIFI module, thereby improving the model inference speed while enhancing the feature extraction capability, especially the improvement effect on the accuracy of small target detection. Secondly, in the neck network, the lightweight dynamic upsampler DySample is used to replace the traditional nn.UPSample, which significantly improves the performance of the model without almost increasing the complexity of the model. Adding a layer of SE attention mechanism to the small feature layer of the original YOLO11 neck network further improves the detection accuracy of small targets. Finally, the head network is responsible for target classification and bounding box regression tasks. By further processing the feature map passed by the neck network, it outputs the category prediction and position prediction results of the target object.

[0072] In summary, the LCSD-YOLO network model takes YOLO11 network as the basic network, and the YOLO11 network includes a backbone network, a neck network and a head network. The backbone network is replaced by PP-LCNet (a lightweight CPU network based on MKLDNN acceleration strategy), and the parameters of the backbone network are as shown in the table. Figure 4 As shown in the table, the down-sampling convolution in the replaced backbone network is replaced by the SDDS module (a depth separable down-sampling convolution that integrates spatial depth conversion), the last layer in the replaced backbone network is replaced by the attention-based internal scale feature interaction module AIFI, the original YOLO11 neck network upsampler nn.UPSample is replaced by the more lightweight dynamic upsampler DySample, and an SE attention mechanism is added to the small feature layer of the original YOLO11 neck network.

[0073] Finally, in this embodiment, the original loss function can be replaced by using a non-monotonic focusing mechanism, which measures the target's WIoUv3 by evaluating the outlying degree of the anchor box, to avoid the precision impact caused by the high or low quality of the data set.

[0074] WIoUv1 introduces distance as an indicator of attention. When the target box and the prediction box overlap within a certain range, the penalty of reducing the geometric measure enables the model to have better generalization ability. The formula for calculating WIoUv1 is as follows:

[0075] L WIoUv1 =R WIoU ×L IoU

[0076]

[0077] L IoU =1-IoU

[0078] In the above formula, represents the x-axis position of the center of the real box, represents the y-axis position of the center of the real box, cx represents the x-axis position of the center of the prediction box, cy represents the y-axis position of the center of the prediction box, w and c h represent the height and width of the minimum bounding box formed by the Prediction Box and the Real Box.

[0079] A non-monotonic focusing coefficient r is constructed using β and applied to WIoUv1 to obtain WIoUv3 with dynamic non-monotonic FM. By using a reasonable gradient gain allocation strategy, the weight of high-quality and low-quality boxes is reduced, so that the model focuses on the average quality samples, thereby improving the performance. The expression for calculating WIoUv1 is as follows:

[0080] L WIoU =r×L WIoUv1

[0081]

[0082] In the above formula, L WIoU represents the target recognition loss function, r represents the non-monotonic focusing coefficient, L WIoUv1 Indicates the indicator function of the degree of overlap between the predicted box and the true box, β represents WIoUv3 with dynamic non-monotonic FM applied to WIoUv1, α and δ represent variable hyperparameters, represents the monotonic focusing coefficient, L IoU It represents the intersection ratio loss function between the predicted box and the real box, and IoU represents the intersection ratio between the predicted box and the real box.

[0083] S300, performing target recognition detection on the preprocessed image dataset based on the LCSD-YOLO network model to obtain a target recognition detection result.

[0084] Specifically, the preprocessed image dataset is input into the LCSD-YOLO network model; the backbone network module based on the LCSD-YOLO network model performs feature recognition processing on the preprocessed image dataset to obtain target image features; the neck network based on the LCSD-YOLO network model performs multi-scale feature fusion processing on the target image features to obtain fused target image features; the head network based on the LCSD-YOLO network model decodes and predicts the fused target image features to obtain target recognition detection results.

[0085] First, we used a self-made dataset of construction worker hardhat images labeled with human and hardhat categories as the model training set. This dataset was divided into a training set and a test set. The training set consisted of 3,000 images, and the validation set consisted of 1,000 images. We then imported the improved LCSD-YOLO network model into a pre-configured deep learning environment, modified the network model's parameter configuration file based on the actual environment, and then trained the model using the processed dataset. We trained the model for 400 epochs, with 64 images per epoch. During training, we monitored the training log in real time using Wandb, and saved the weight files after training was complete.

[0086] After model training is complete, a corresponding weight file is generated. Select the best.pt weight file, the image to be tested, and the label, and import it into the test program. After the program executes, it will output images of the test results. Analyze these images to evaluate whether the experiment meets the expected requirements.

[0087] Furthermore, for the backbone network module, the preprocessed image dataset is input into the backbone network module of the LCSD-YOLO network model; based on the first convolution module of the backbone network module, channel adjustment and feature extraction are performed on the preprocessed image dataset to obtain a preliminary target feature image; based on the SDDS module of the backbone network module, convolution and spatial channel dimension conversion are performed on the preliminary target feature image to obtain a first intermediate target feature image; based on the depthwise separable convolution module of the backbone network module, depthwise separable convolution is performed on the intermediate target feature image to obtain a second intermediate target feature image; based on the attention-based internal scale feature interaction module of the backbone network module, multi-scale feature interaction redundancy processing is performed on the second intermediate target feature image to obtain target image features.

[0088] Furthermore, for the SDDS module in the backbone network module, the preliminary target feature image is input into the SDDS module of the backbone network module, and the SDDS module includes a depth-by-depth convolution module, a point-by-point convolution module and a spatial depth conversion module; based on the depth-by-depth convolution module of the SDDS module, the preliminary target feature image is subjected to depth-by-depth convolution processing to obtain a first convolved target feature image; based on the point-by-point convolution module of the SDDS module, the target feature image after the first convolution is subjected to channel aggregation processing to obtain a second convolved target feature image; based on the spatial depth conversion module of the SDDS module, the target feature image after the second convolution is connected along the channel dimension to obtain a first intermediate target feature image.

[0089] For the neck network, the target image features are input into the neck network of the LCSD-YOLO network model; the DySample module based on the neck network performs feature reconstruction processing on the target image features to obtain reconstructed target image features; the Concat module based on the neck network performs feature splicing processing on the reconstructed target image features to obtain spliced ​​target image features; the C3K2 module based on the neck network performs multi-level spatial feature extraction processing on the spliced ​​target image features to obtain deeper target image features; the SE module based on the neck network performs channel feature weight adjustment processing on the deeper target image features to obtain adjusted target image features; the convolution module based on the neck network performs convolution processing on the adjusted target image features to output the fused target image features.

[0090] In this embodiment, the image of the high-altitude workers collected by the ground camera looking up has a low proportion of target pixels (usually less than 0.1% of the image area). After the improvement, the input image first undergoes a 3×3 convolution for channel adjustment and feature extraction, and then the feature map is downsampled by the depth-separable downsampling convolution that integrates the spatial depth conversion through the SDDS layer. The depth-separable convolution consists of depth-wise convolution and point-wise convolution. The depth-wise convolution is used to extract spatial features, and the point-by-point convolution is used to extract channel features. The depth-wise separable convolution groups the convolutions in the feature dimension, performs independent depth-wise convolution on each channel, and uses a 1x1 convolution to aggregate all channels before output, such as Figure 5 as well as Figure 6 As shown, due to the need for grouped convolution, the number of input channels should always be a multiple of the number of output channels. In order to accelerate the model's ability to extract features, the present invention integrates spatial depth conversion, and the feature map X obtained by connecting the depth-separable convolution is shown as Figure 7 As shown, the module defaults to 2 and obtains 4 subgraphs f 0,0 、f 1,0 、f 0,1 With f 1,1 Each has shape (S / 2, S / 2, C1) and downsamples X by a factor of 2. Next, these sub-feature maps are concatenated along the channel dimension to obtain a feature map X′ whose spatial dimension is reduced by a scaling factor and whose channel dimension is increased by a scaling factor of 2. In other words, SPD transforms the feature map X(S, S, C1) into an intermediate feature map X′(S / 2, S / 2, C1) 2C1), and batch normalization can further improve the quality of feature maps, making it easier for subsequent operations to capture effective patterns. At the same time, combined with the Hardswish activation function, it can ensure that the transformed feature maps still maintain good statistical characteristics, thereby maximizing the efficiency of the overall architecture. The attention-based intra-scale feature interaction module AIFI reduces the computational cost of feature extraction by only extracting features on the third depth-separable convolution module and reducing the redundancy of multi-level feature extraction. It avoids the situation where high-level features extracted from low-level features contain too much rich semantic information about the object and redundant feature interaction on the cascaded multi-scale features. By using a single-scale Transformer encoder to perform intra-scale interaction only on the third depth-separable convolution module, the computational cost is further reduced. The feature map then passes through the dynamic upsampling layer DySample, which samples the essential points of the upsampling to generate content-aware sampling points. The point offset is generated by linear projection, and the point value is resampled using PyTorch's grid_sample function. The upsampling of the feature map is gradually optimized by controlling the initial sampling position, adjusting the offset range, and grouping the upsampling process. In addition, the free convolution kernel is enabled only in the fourth C3K2 layer, which improves the extraction capability of the feature map without increasing the amount of invalid calculation too much. The optimization effect of the free convolution kernel on feature extraction in the low layer is not obvious.

[0091] like Figure 8 As shown in Figure 2, the DySample dynamic upsampling module assumes l input features, upsampled features, generated offsets, original grids, and siqmoid activation functions, denoted by X, X′, 0, G, and σ respectively. Figure 8 (a) in the figure indicates that the sampling set is generated by the sampling point generator, and the input features are resampled by the grid sample function. Figure 8 (b) in the figure shows that in the generator, the sample set is the sum of the generated offset and the original grid position. The upper box shows the version with a "static range factor," where the offset is generated by a linear layer. The lower box depicts the version with a "dynamic range factor," where the range factor is first generated and then used to modulate the offset.

[0092] In summary, the present invention uses a lightweight backbone network PP-LCNet to replace the original yolo11n backbone network (a lightweight network using depth-separable convolution) to reduce model complexity and reduce computing resource requirements, and replaces the original downsampling convolution of the PP-LCNet backbone network with a depth-separable downsampling convolution SDDS that integrates spatial depth conversion, reducing the loss of fine-grained information and improving the learning efficiency of feature representation, significantly improving the problem of insufficient small target detection capability caused by the impaired feature expression capability of the direct compression model. At the same time, the use of Hardswish activation function and batch normalization greatly improves the inference speed of the model, meeting the requirements of real-time detection at the edge. Enabling the SE mechanism enhances the model's ability to extract target features by dynamically adjusting channel weights, while the AIFI module only interacts within the scale on the third depth-separable convolution module, further reducing the computational cost while improving the model's feature extraction capability, and effectively solving the problem of insufficient small target detection accuracy caused by the damage to the feature pyramid of the lightweight model. Using the dynamic upsampling layer DySample, the performance of the model is further improved without changing the complexity of the original model. This technical solution has the advantages of high computational efficiency, strong feature extraction capability, and flexible scalability. It is very suitable for the safety of construction site personnel and meets the requirements of high precision, real-time performance, and low latency in practical applications. The recognition results are as follows: Figure 9 shown.

[0093] Reference Figure 2 , an object detection system based on LCSD-YOLO network, including:

[0094] The first module 201 is used to obtain target image data and perform data preprocessing to obtain a preprocessed image data set;

[0095] The second module 202 is used to introduce the PP-LCNet network architecture, the SDDS module and the attention-based internal scale feature interaction module based on the YOLO11 network model to build an LCSD-YOLO network model;

[0096] The third module 203 is used to perform target recognition detection on the preprocessed image dataset based on the LCSD-YOLO network model to obtain a target recognition detection result.

[0097] The contents of the above method embodiments are all applicable to the present system embodiments. The functions specifically implemented by the present system embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0098] The above is a specific description of the preferred implementation of the present invention, but the invention is not limited to the embodiments. Those skilled in the art can make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

1. A target detection method based on LCSD-YOLO network, characterized in that: The following steps are involved: Acquire target image data and perform data preprocessing to obtain a preprocessed image data set; Based on the YOLO11 network model, the PP-LCNet network architecture, SDDS module and attention-based internal scale feature interaction module are introduced to build the LCSD-YOLO network model; The LCSD-YOLO network model specifically includes a backbone network module, a neck network, and a head network, wherein the backbone network module, the neck network, and the head network are connected in sequence, wherein: The backbone network module adopts the PP-LCNet network architecture, and the backbone network module includes a first convolution module, an SDDS module, a depth-separable convolution module and an attention-based internal scale feature interaction module. The SDDS module includes a first SDDS module, a second SDDS module and a third SDDS module. The depth-separable convolution module includes a first depth-separable convolution module, a second depth-separable convolution module and a third depth-separable convolution module. The SDDS module includes a depth-by-depth convolution module, a point-by-point convolution module and a spatial depth conversion module; The neck network includes a DySample module, a Concat module, a C3K2 module, an SE module and a convolution module, the DySample module includes a first DySample module and a second DySample module, the Concat module includes a first Concat module, a second Concat module, a third Concat module and a fourth Concat module, the first C3K2 module, the second C3K2 module, the third C3K2 module and the fourth C3K2 module, and the convolution module includes a second convolution module and a third convolution module; The head network includes a first detection head, a second detection head and a third detection head; The LCSD-YOLO network model constructs a target recognition loss function by introducing a non-monotonic focusing mechanism. The expression of the target recognition loss function is as follows: In the above formula, represents the target recognition loss function, represents the non-monotonic focusing coefficient, An indicator function that represents the degree of overlap between the predicted box and the true box, Indicates application to With dynamic non-monotonic FM , and represents a variable hyperparameter, represents the monotonic focusing coefficient, Represents the intersection ratio loss function of the predicted box and the real box, Represents the intersection ratio of the predicted box and the real box; Based on the LCSD-YOLO network model, target recognition detection is performed on the preprocessed image dataset to obtain the target recognition detection results.

2. The target detection method based on the LCSD-YOLO network according to claim 1, wherein The step of acquiring target image data and performing data preprocessing to obtain a preprocessed image data set specifically includes: Capture images of the target object through a camera to obtain a target image dataset; Name the target image dataset according to the format of the Pascal VOC dataset and create target image data files with three folders: Annotations, ImageSets, and JPEGImages; Perform data labeling on the target image data file to obtain a preprocessed image dataset.

3. The target detection method based on the LCSD-YOLO network according to claim 2, characterized in that: The step of performing target recognition detection on the preprocessed image dataset based on the LCSD-YOLO network model to obtain the target recognition detection result specifically includes: Input the preprocessed image dataset into the LCSD-YOLO network model; Based on the backbone network module of the LCSD-YOLO network model, feature recognition processing is performed on the preprocessed image dataset to obtain the target image features; Based on the neck network of the LCSD-YOLO network model, multi-scale feature fusion processing is performed on the target image features to obtain the fused target image features; Based on the head network of the LCSD-YOLO network model, the fused target image features are decoded and predicted to obtain the target recognition detection results.

4. The target detection method based on the LCSD-YOLO network according to claim 3, wherein: The backbone network module based on the LCSD-YOLO network model performs feature recognition processing on the preprocessed image data set to obtain target image features, which specifically includes: Input the preprocessed image dataset into the backbone network module of the LCSD-YOLO network model; Based on the first convolution module of the backbone network module, channel adjustment and feature extraction are performed on the preprocessed image dataset to obtain a preliminary target feature image; The SDDS module based on the backbone network module performs convolution and spatial channel dimension conversion on the preliminary target feature image to obtain the first intermediate target feature image; Based on the depthwise separable convolution module of the backbone network module, a depthwise separable convolution process is performed on the intermediate target feature image to obtain a second intermediate target feature image; The attention-based internal scale feature interaction module based on the backbone network module performs multi-scale feature interaction redundant processing on the second intermediate target feature image to obtain the target image features.

5. The target detection method based on the LCSD-YOLO network according to claim 4, characterized in that: The SDDS module based on the backbone network module performs convolution and spatial channel dimension conversion on the preliminary target feature image to obtain the first intermediate target feature image, which specifically includes: Input the preliminary target feature image into the SDDS module of the backbone network module, wherein the SDDS module includes a depth-wise convolution module, a point-wise convolution module, and a spatial depth conversion module; Based on the depth-wise convolution module of the SDDS module, the preliminary target feature image is subjected to depth-wise convolution processing to obtain the target feature image after the first convolution; Based on the point-by-point convolution module of the SDDS module, channel aggregation processing is performed on the target feature image after the first convolution to obtain the target feature image after the second convolution; Based on the spatial depth conversion module of the SDDS module, the target feature image after the second convolution is connected along the channel dimension to obtain the first intermediate target feature image.

6. The target detection method based on the LCSD-YOLO network according to claim 5, characterized in that: The neck network based on the LCSD-YOLO network model performs multi-scale feature fusion processing on the target image features to obtain the fused target image features. This step specifically includes: Input the target image features into the neck network of the LCSD-YOLO network model; The DySample module based on the neck network performs feature reconstruction on the target image features to obtain the reconstructed target image features; The Concat module based on the neck network performs feature splicing on the reconstructed target image features to obtain the spliced ​​target image features; The C3K2 module based on the neck network performs multi-level spatial feature extraction on the spliced ​​target image features to obtain deeper target image features; Based on the SE module of the neck network, the channel feature weight adjustment processing is performed on the deeper target image features to obtain the adjusted target image features; The convolution module based on the neck network performs convolution processing on the adjusted target image features and outputs the fused target image features.

7. A target detection system based on LCSD-YOLO network, characterized in that: Includes the following modules: The first module is used to obtain target image data and perform data preprocessing to obtain a preprocessed image data set; The second module is used to build the LCSD-YOLO network model based on the YOLO11 network model by introducing the PP-LCNet network architecture, the SDDS module, and the attention-based internal scale feature interaction module; The LCSD-YOLO network model specifically includes a backbone network module, a neck network, and a head network, wherein the backbone network module, the neck network, and the head network are connected in sequence, wherein: The backbone network module adopts the PP-LCNet network architecture, and the backbone network module includes a first convolution module, an SDDS module, a depth-separable convolution module and an attention-based internal scale feature interaction module. The SDDS module includes a first SDDS module, a second SDDS module and a third SDDS module. The depth-separable convolution module includes a first depth-separable convolution module, a second depth-separable convolution module and a third depth-separable convolution module. The SDDS module includes a depth-by-depth convolution module, a point-by-point convolution module and a spatial depth conversion module; The neck network includes a DySample module, a Concat module, a C3K2 module, an SE module and a convolution module, the DySample module includes a first DySample module and a second DySample module, the Concat module includes a first Concat module, a second Concat module, a third Concat module and a fourth Concat module, the first C3K2 module, the second C3K2 module, the third C3K2 module and the fourth C3K2 module, and the convolution module includes a second convolution module and a third convolution module; The head network includes a first detection head, a second detection head and a third detection head; The LCSD-YOLO network model constructs a target recognition loss function by introducing a non-monotonic focusing mechanism. The expression of the target recognition loss function is as follows: In the above formula, represents the target recognition loss function, represents the non-monotonic focusing coefficient, An indicator function that represents the degree of overlap between the predicted box and the true box, Indicates application to With dynamic non-monotonic FM , and represents a variable hyperparameter, represents the monotonic focusing coefficient, Represents the intersection ratio loss function of the predicted box and the real box, Represents the intersection ratio of the predicted box and the real box; The third module is used to perform target recognition detection on the preprocessed image dataset based on the LCSD-YOLO network model to obtain the target recognition detection results.

Images