Multi-scale tower crane component identification method based on YOLO model

Abstract

The invention provides a multi-scale tower crane component identification method based on a YOLO model. The method comprises the following steps: acquiring tower crane component image data; inputting the tower crane component image data into the SLMYOLO model and processing to generate an enhanced data set according with model input; a backbone network of the SLMYOLO model comprises a plurality of convolutional layers and a C2fSCConv layer, and effective feature layers are extracted through neck networks of FPN and PAN structures; and finally, the physical position and category information are output through the detection head. According to the invention, through the improved network structure and algorithm, the accuracy and efficiency of tower crane part identification are significantly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of computer vision and image target detection, and in particular to a multi-scale tower crane component recognition method based on a YOLO model. Background Art

[0002] With the rapid development of computer technology, its impact in various industries is becoming increasingly significant. Among them, target detection technology, as an important direction in the field of computer vision, has been widely used in daily life, including intelligent video surveillance, traffic monitoring, robot navigation, aerospace and other fields. In the construction industry, target detection technology also plays an important role, such as safety helmet detection, dangerous area intrusion monitoring and face recognition punching. Tower cranes are important equipment widely used in construction sites, and safety in construction is of vital importance. However, the current maintenance of tower crane components still relies mainly on manual work. By introducing target detection technology, it can help maintenance personnel quickly locate key components and improve maintenance efficiency. Due to the complex site environment, tower crane components are large and often have oil stains on the surface. In addition, when taking pictures, they are affected by factors such as angle, distance and light, which easily produce shadows, complex backgrounds and unclear target features. These have brought many challenges to the target detection task. Therefore, in the detection of tower crane components, how to improve the accuracy and robustness of target detection in complex environments has become an important problem that needs to be solved urgently.

[0003] Object detection technology based on deep learning is mainly divided into two camps: two-stage and one-stage. Two-stage algorithms, such as the R-CNN series, generate candidate boxes and classify them finely. Although they have excellent accuracy, the redundant candidate boxes lead to large memory consumption. In contrast, one-stage algorithms, especially the YOLO series, have significantly improved the detection speed and met the real-time requirements by directly converting the detection task into a regression problem since the advent of YOLOv1 in 2016. The YOLO series is constantly iterating, from v1 to the latest v8, and each version marks a leap in single-stage detection technology. As the current SOTA model, YOLOv8 not only inherits the efficiency and modular design of its predecessor, but also incorporates a number of innovations to further improve detection accuracy and flexibility. It can process high-resolution images in real time, accurately identify a variety of targets, and can run smoothly even on low-end devices with limited resources. In view of these advantages of YOLOv8, the present invention intends to use it as a basic model to perform efficient and accurate recognition of tower crane components. Summary of the invention

[0004] The purpose of the present invention is to propose a multi-scale tower crane component recognition method based on the YOLO model to address the current shortcomings.

[0005] The present invention adopts the following technical solution:

[0006] A multi-scale tower crane component recognition method based on a YOLO model, the method comprising the following steps:

[0007] S1: Acquire image data of tower crane components; this step includes:

[0008] S1.1: Collect image data of tower crane components;

[0009] S1.2: Data augmentation of images by random cropping and random flipping techniques;

[0010] S1.3: Annotate all enhanced images and use the annotated images as a tower crane component dataset;

[0011] S1.4: Divide the tower crane component dataset into a training set and a test set in a ratio of 9:1;

[0012] S2: Input the image data of the tower crane components into the SLM_YOLO model and process it to generate an enhanced data set that conforms to the model input; this step includes:

[0013] S2.1: Input the images in the training set into the SLM_YOLO model;

[0014] S2.2: Preprocess the image; the preprocessing operation includes: adjusting the image size to scale the image length and width in equal proportion. If the target size is not reached, fill the pixels at the edge of the image and finally adjust the image size to 640x640; after the image size is adjusted, normalize the image pixel value to the interval [0,1],

[0015] S2.3: Perform Mosaic data enhancement operation on all preprocessed images to obtain enhanced data sets;

[0016] S3: extracting features from the enhanced data set through the backbone network of the SLM_YOLO model to obtain a feature map; wherein the backbone network structure includes five convolutional layers and four C2f_SCConv layers; the enhanced data set is subjected to feature extraction through the backbone network and then outputs a feature map;

[0017] S4: Input the feature map output in the previous step into the neck network part of the YOLO model to extract the effective feature layer; the neck network part adopts a structure combining FPN and PAN to achieve mutual fusion of shallow features and deep features, and the neck network part extracts three effective feature layers in total;

[0018] S5: input the three effective feature layers into the detection head to complete target detection and recognition; the detection head is composed of two branches, and the two branches are used to extract the location information and category information of the object respectively;

[0019] S6: Use cross-validation to optimize model parameters and use a composite loss function to optimize the prediction accuracy of both location and category.

[0020] S7: Use the test set to evaluate the performance of the SLM_YOLO model; and adjust the model's hyperparameters in the process to achieve the best performance of the model;

[0021] S8: Deploy and apply the trained and verified SLM_YOLO model in actual usage scenarios to complete the automatic recognition and classification of tower base components.

[0022] Furthermore, the C2f_SCConv layer in the backbone network includes a convolution module, a Split module, two Bottleneck_SCConv layers and a splicing layer; the image in the enhanced data set enters the C2f_SCConv layer for deep processing after feature extraction in the convolution layer, and the specific processing method is as follows:

[0023] S3.1: Receive the intermediate feature map output by the previous layer of the C2f_SCConv layer through the convolution module, and perform further convolution processing on the intermediate feature map;

[0024] S3.2: The intermediate feature map after the convolution processing in the previous step is divided into two parts through the Split module, one part is passed to the final concatenation layer, and the other part is sent to two Bottleneck_SCConv layers for processing; each Bottleneck_SCConv layer contains a normal convolution and a self-calibration convolution;

[0025] S3.3: When the first Bottleneck_SCConv layer processes the intermediate feature map, it continuously outputs the processing result through a normal convolution and a self-calibration convolution; the output processing result of the first Bottleneck_SCConv layer is divided into two parts, one part enters the next Bottleneck_SCConv layer for processing, and the other part enters the final splicing layer; the second Bottleneck_SCConv layer also continuously outputs the processing result through a normal convolution and a self-calibration convolution; the output processing result of the second Bottleneck_SCConv layer directly enters the final splicing layer.

[0026] Furthermore, the specific implementation process of the self-calibration convolution is as follows:

[0027] S3.3.1: Average pooling: Let the input to the self-calibration convolution intermediate feature map be X1, the filter size be r×r, and the stride be r; then:

[0028] T1=AvgPoolr (X1);

[0029] Where T1 is the downsampled intermediate feature map obtained by applying an average pooling function of size r×r and stride r; AvgPool r is the average pooling function,

[0030] S3.3.2: Feature enhancement and upsampling:

[0031] X'1=Up(F2(T1))=Up(T1*K2);

[0032] Where X′1 is the intermediate feature map obtained by applying K2 operation to T1; Up() is a bilinear interpolation operator, F2(T1)=T1*K2, where K2 includes an average pooling operation and a convolution operation, and F2(T1) means applying K2 to T1;

[0033] S3.3.3: Self-calibration operation:

[0034] Y′1=F3(X1)·σ(X1+X′1);

[0035] Where Y′1 is the self-calibration output; F3(X1)=X1*K3, where K3 represents the convolution operation, F3(X1) represents applying the convolution K3 to X1, and σ is the sigmoid activation function;

[0036] S3.3.4: Final convolution output:

[0037] Y1=F4(Y′1)=Y′1*K4

[0038] Among them, Y1 is the final output of the self-calibration convolution, K4 represents the final convolution operation, and F4(Y′1) represents the application of convolution K4 to Y′1.

[0039] Furthermore, the FPN structure in the neck network includes an SPPF_LSKA layer, two upsampling layers, two concatenation layers and two C2f_SCConv layers; the PAN structure includes two convolution layers, two concatenation layers and two C2f_SCConv layers.

[0040] Furthermore, the feature map finally outputted in the backbone network enters the SPPF_LSKA layer; the SPPF_LSKA layer includes two convolutional layers, three pooling layers, an LSKA layer and a concatenation layer; the processing process of the feature map by the SPPF_LSKA layer is as follows:

[0041] S4.1: The feature map enters the convolution layer in the SPPF_LSKA layer for preliminary feature extraction; the features after convolution layer processing Figure 1Some are sent to the concatenation layer, and the other part goes to the subsequent pooling layer;

[0042] S4.2: The part sent to the pooling layer in the previous step enters three consecutive maximum pooling layers. Each pooling layer captures features of different scales through pooling windows of different sizes. Part of the output of the first and second pooling layers directly enters the concatenation layer, and the other part enters the subsequent pooling layer. The output of the third pooling layer enters the LSKA layer.

[0043] S4.3: The part entering the LSKA layer first passes through two depthwise separable convolutions and then enters two dilated convolution layers, thereby increasing the spatial coverage of the feature map; the output of the LSKA layer enters the subsequent concatenation layer;

[0044] S4.4: The concatenation layer integrates and fuses the feature maps received in all the previous steps, and inputs the integrated feature maps into the final convolutional layer to weight the feature maps to generate weighted feature maps.

[0045] Furthermore, after generating the weighted feature map, the SPPF_LSKA layer sends the weighted feature map to the subsequent FPN and PAN structures of the neck network to extract three effective feature layers.

[0046] Furthermore, in step S5, the branch for extracting the object's position information extracts the object's position information through a 3x3 convolution, a MultiSEAM layer, a 1x1 convolution and Bbox.loss; the branch for extracting the object's category information extracts the object's category information through a 3x3 convolution, a MultiSEAM layer, a 1x1 convolution and Cls.loss.

[0047] Beneficial effects achieved by the present invention:

[0048] The present invention can accurately and efficiently identify tower base components through the SLM_YOLO model; by setting the Bottleneck_SCConv layer in the C2f_SCConv layer of the model, the model can capture more complex patterns and details, and further refine and enhance features; the weighted feature map is obtained through the SPPF_LSKA layer, which highlights important features and suppresses unimportant features, greatly improving the ability of the SPPF_LSKA layer to aggregate features at multiple scales; the MultiSEAM layer in the detection head can better help the model extract feature information. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The present invention can be further understood from the following description in conjunction with the accompanying drawings. The components in the figures are not necessarily drawn to scale, but the emphasis is placed on illustrating the principles of the embodiments. In different views, the same reference numerals designate corresponding parts.

[0050] Figure 1 It is a schematic diagram of the multi-scale tower crane component recognition method based on the YOLO model of the present invention.

[0051] Figure 2 Schematic diagram of the network structure of the SLM_YOLO model of the present invention.

[0052] Figure 3 Schematic diagram of the C2f_SCConv layer structure of the present invention.

[0053] Figure 4 It is a schematic diagram of the SPPF_LSKA layer structure of the present invention.

[0054] Figure 5 It is a schematic diagram of the structure of the detection head of the present invention. DETAILED DESCRIPTION

[0055] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with its embodiments; it should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention; for those skilled in the art, other systems, methods and / or features of the present embodiment will become apparent after reviewing the following detailed description; it is intended that all such additional systems, methods, features and advantages are included in this specification; included within the scope of the present invention and protected by the appended claims; additional features of the disclosed embodiments are described in the following detailed description, and these features will be apparent from the following detailed description.

[0056] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right" and the like indicate directions or positional relationships based on the directions or positional relationships shown in the drawings, it is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or component referred to must have a specific direction, be constructed and operated in a specific direction. Therefore, the terms describing the positional relationship in the drawings are only used for illustrative purposes and cannot be understood as limitations on this patent. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0057] Embodiment 1:

[0058] like Figure 1 , Figure 2 As shown, this embodiment provides a multi-scale tower crane component recognition method based on the YOLO model, and the method includes the following steps:

[0059] S1: Acquire image data of tower crane components; this step includes:

[0060] S1.1: Collect image data of tower crane components;

[0061] S1.2: Data augmentation of images by random cropping and random flipping techniques;

[0062] S1.3: Annotate all enhanced images and use the annotated images as a tower crane component dataset;

[0063] S1.4: Divide the tower crane component dataset into a training set and a test set in a ratio of 9:1;

[0064] S2: Input the image data of the tower crane components into the SLM_YOLO model and process it to generate an enhanced data set that conforms to the model input; this step includes:

[0065] S2.1: Input the images in the training set into the SLM_YOLO model;

[0066] S2.2: Preprocess the image; the preprocessing operation includes: adjusting the image size to scale the image length and width in equal proportion. If the target size is not reached, fill the pixels at the edge of the image and finally adjust the image size to 640x640; after the image size is adjusted, normalize the image pixel value to the interval [0,1],

[0067] S2.3: Perform Mosaic data enhancement operation on all preprocessed images to obtain enhanced data sets;

[0068] S3: Extract features of the enhanced data set through the backbone network part of the SLM_YOLO model to obtain a feature map; wherein the structure of the backbone network includes five convolutional layers and four C2f_SCConv layers; the enhanced data set is subjected to feature extraction through the backbone network and a feature map is output, and the feature map is sent to the SPPF_LSKA layer for processing to obtain a weighted feature map;

[0069] S4: Input the feature map output in the previous step into the neck network part of the SLM_YOLO model to extract the effective feature layer; the neck network part adopts a structure combining FPN and PAN to achieve the mutual fusion of shallow features and deep features, and the neck network part extracts a total of three effective feature layers;

[0070] S5: input the three effective feature layers into the detection head to complete target detection and recognition; the detection head is composed of two branches, and the two branches are used to extract the location information and category information of the object respectively;

[0071] S6: Use cross-validation to optimize model parameters and use a composite loss function to optimize the prediction accuracy of both location and category.

[0072] S7: Use the test set to evaluate the performance of the SLM_YOLO model; and adjust the model's hyperparameters in the process to achieve the best performance of the model;

[0073] S8: Deploy and apply the trained and verified SLM_YOLO model in actual usage scenarios to automatically identify and classify tower base components;

[0074] Further, such as Figure 3 As shown, the C2f_SCConv layer in the backbone network includes a convolution module, a Split module, two Bottleneck_SCConv layers and a splicing layer; the image in the enhanced dataset enters the C2f_SCConv layer for deep processing after feature extraction in the convolution layer, and the specific processing method is as follows:

[0075] S3.1: Receive the intermediate feature map output by the previous layer of the C2f_SCConv layer through the convolution module, and perform further convolution processing on the intermediate feature map;

[0076] S3.2: The intermediate feature map after the convolution processing in the previous step is divided into two parts through the Split module, one part is passed to the final concatenation layer, and the other part is sent to two Bottleneck_SCConv layers for processing; each Bottleneck_SCConv layer contains a normal convolution and a self-calibration convolution;

[0077] S3.3: When the first Bottleneck_SCConv layer processes the intermediate feature map, it continuously outputs the processing result through a common convolution and a self-calibration convolution; the output processing result of the first Bottleneck_SCConv layer is divided into two parts, one part enters the next Bottleneck_SCConv layer for processing, and the other part enters the final concatenation layer; the second Bottleneck_SCConv layer also continuously outputs the processing result through a common convolution and a self-calibration convolution; the output processing result of the second Bottleneck_SCConv layer directly enters the final concatenation layer;

[0078] The specific implementation process of the self-calibration convolution is as follows:

[0079] S3.3.1: Average pooling: Let the input to the self-calibration convolution intermediate feature map be X1, the filter size be r×r, and the stride be r; then:

[0080] T1=AvgPool r (X1);

[0081] Where T1 is the downsampled intermediate feature map obtained by applying an average pooling function of size r×r and stride r; AvgPool r is the average pooling function,

[0082] S3.3.2: Feature enhancement and upsampling:

[0083] X'1=Up(F2(T1))=Up(T1*K2);

[0084] Where X′1 is the intermediate feature map obtained by applying K2 operation to T1; Up() is a bilinear interpolation operator, F2(T1)=T1*K2, where K2 includes an average pooling operation and a convolution operation, and F2(T1) means applying K2 to T1;

[0085] S3.3.3: Self-calibration operation:

[0086] Y′1=F3(X1)·σ(X1+X′1);

[0087] Where Y′1 is the self-calibration output; F3(X1)=X1*K3, where K3 represents the convolution operation, F3(X1) represents applying the convolution K3 to X1, and σ is the sigmoid activation function;

[0088] S3.3.4: Final convolution output:

[0089] Y1=F4(Y′1)=Y′1*K4

[0090] Among them, Y1 is the final output of the self-calibration convolution, K4 represents the final convolution operation, and F4(Y′1) represents the application of convolution K4 to Y′1;

[0091] In this scheme, the Bottleneck in the basic model is improved to Bottleneck_SCConv, which is a self-calibrated convolution. Unlike the standard convolution that uses a small-size kernel to fuse spatial and channel information at the same time, SCConv can adaptively build long-range spatial and channel dependencies around each spatial position; this feature of SCConv can help CNN generate more discriminative feature expressions, expand the field of view of each convolution kernel, and thus enrich the output features; ultimately, the Bottleneck_SCConv layer has better feature extraction capabilities and stronger feature expression capabilities;

[0092] Furthermore, the FPN structure in the neck network includes an SPPF_LSKA layer, two upsampling layers, two concatenation layers, and two C2f_SCConv layers; the PAN structure includes two convolution layers, two concatenation layers, and two C2f_SCConv layers;

[0093] Further, such as Figure 4 As shown, the feature map finally output in the backbone network enters the SPPF_LSKA layer; the SPPF_LSKA layer includes two convolutional layers, three pooling layers, an LSKA layer and a splicing layer; the processing process of the SPPF_LSKA layer for the feature map is as follows:

[0094] S4.1: The feature map enters the convolution layer in the SPPF_LSKA layer for preliminary feature extraction; the features after convolution layer processing Figure 1 Some are sent to the concatenation layer, and the other part goes to the subsequent pooling layer;

[0095] S4.2: The part sent to the pooling layer in the previous step enters three consecutive maximum pooling layers. Each pooling layer captures features of different scales through pooling windows of different sizes. Part of the output of the first and second pooling layers directly enters the concatenation layer, and the other part enters the subsequent pooling layer. The output of the third pooling layer enters the LSKA layer.

[0096] S4.3: The part entering the LSKA layer first passes through two depthwise separable convolutions and then enters two dilated convolution layers, thereby increasing the spatial coverage of the feature map; the output of the LSKA layer enters the subsequent concatenation layer;

[0097] S4.4: The concatenation layer integrates and fuses the feature maps received in all the previous steps, and inputs the integrated feature maps into the final convolution layer to weight the feature maps to generate weighted feature maps;

[0098] In this solution, the SPPF_LSKA layer introduces the LSKA layer into the original SPPF layer. The attention mechanism of the LSKA layer captures the extensive contextual information of the image by using large and separable convolution kernels and spatial dilation convolutions, generates an attention map, and weights the original features through the attention map, thereby enhancing the network's attention to important features and improving the performance of the model; the SPPF_LSKA layer can improve the model's ability to capture multi-scale features and context, improve feature selectivity, and reduce the computational complexity of the model;

[0099] Furthermore, after generating the weighted feature map, the SPPF_LSKA layer sends the weighted feature map to the subsequent FPN and PAN structures of the neck network to extract three effective feature layers;

[0100] Furthermore, in step S4, the three extracted effective feature layers are defined as: feat1 = (80, 80, 256*w), feat2 = (40, 40, 512*w), feat3 = (20, 20, 512*w*r); wherein w and r are model adjustment coefficients;

[0101] Further, such as Figure 5 As shown, in step S5, the branch for extracting the object's position information extracts the object's position information through a 3x3 convolution layer, a MultiSEAM layer, a 1x1 convolution layer, and Bbox.loss; the branch for extracting the object's category information extracts the object's category information through a 3x3 convolution layer, a MultiSEAM layer, a 1x1 convolution layer, and Cls.loss;

[0102] Specifically, the Bbox.loss is the bounding box loss, which is used to regress the bounding box coordinates of the object, including the center point coordinates, width and height; the Cls.loss is the classification loss, which is used to classify the object category and evaluate the error between the predicted category and the true category;

[0103] Specifically, after entering the detection head, the effective feature layer first undergoes a 3x3 convolution for feature extraction, and then enters the MultiSEAM layer; the MultiSEAM layer divides the effective feature layer into three parts on average, passes through convolution layers with convolution kernel sizes of 3, 5 and 7 respectively, and then adds them to the original effective feature layer and sends them to the pooling layer for average pooling, and the output is the element-by-element product of the feature map and the attention weight; the MultiSEAM layer undergoes a 1x1 convolution layer for feature extraction, and finally completes the recognition of object position information and category information through bounding box loss and classification loss.

[0104] Embodiment 2:

[0105] This embodiment should be understood to include at least all the features of any of the above embodiments, and further improve upon them;

[0106] In this embodiment, the SLM_YOLO model is trained using the training set; the training is iterated 300 times on the server, and the log information of each iteration is systematically collected during the training process, and the log information is visualized and analyzed; the results show that the loss function of the SLM_YOLO model shows a continuous downward trend during the iteration process, indicating that the model shows good convergence during the training process and can reach a stable state in a relatively short time;

[0107] In order to comprehensively evaluate the performance of the SLM_YOLO model, mAP50 and mAP50-95 are used as the main evaluation indicators; in the same experimental environment and dataset, the SLM_YOLO model is subjected to ablation experiments and compared with several other mainstream target detection models. The results of the performance evaluation are shown in the following table:

[0108]

[0109]

[0110] Through ablation experiments, it is found that each individual improvement point improves the performance of the model on the tower crane component dataset; specifically, the improvement of the SCConv part improves the mAP50 value by about 1.5 percentage points, the improvement of the LSKA part improves by about 2.1 percentage points, and the improvement of the MultiSEAM part is the most significant, by about 4.2 percentage points; the SLM_YOLO model combining these three improvement points improves mAP50 by 6.4 percentage points, reaching 91.0%, which is significantly higher than other comparison models; the experimental results show that the SLM_YOLO model shows excellent accuracy and efficiency in the recognition task of tower crane components. This embodiment confirms the superior performance of the SLM_YOLO model improved by integrating SCConv, LSKA and MultiSEAM in handling complex detection tasks.

[0111] The contents disclosed above are only preferred feasible embodiments of the present invention, and do not limit the protection scope of the present invention. Therefore, all equivalent technical changes made using the contents of the present invention specification and drawings are included in the protection scope of the present invention. In addition, the elements therein can be updated as technology develops.

Claims

1. A multi-scale tower crane component recognition method based on the YOLO model, characterized in that: The method comprises the following steps: S1: Acquire image data of tower crane components; this step includes: S1.1: Collect image data of tower crane components; S1.2: Data augmentation of images by random cropping and random flipping techniques; S1.3: Annotate all enhanced images and use the annotated images as a tower crane component dataset; S1.4: Divide the tower crane component dataset into a training set and a test set in a ratio of 9:1; S2: Input the image data of the tower crane components into the SLM_YOLO model and process it to generate an enhanced data set that conforms to the model input; this step includes: S2.1: Input the images in the training set into the SLM_YOLO model; S2.2: Preprocess the image; the preprocessing operation includes: adjusting the image size to scale the image length and width in equal proportion. If the target size is not reached, fill the pixels at the edge of the image and finally adjust the image size to 640x640. After the image size adjustment is completed, the image pixel values ​​are normalized to the interval [0,1]. S2.3: Perform Mosaic data enhancement operation on all preprocessed images to obtain enhanced data sets; S3: extracting features from the enhanced data set through the backbone network of the SLM_YOLO model to obtain a feature map; wherein the backbone network structure includes five convolutional layers and four C2f_SCConv layers; the enhanced data set is subjected to feature extraction through the backbone network and then outputs a feature map; S4: Input the feature map output in the previous step into the neck network part of the YOLO model to extract the effective feature layer; the neck network part adopts a structure combining FPN and PAN to achieve mutual fusion of shallow features and deep features, and the neck network part extracts three effective feature layers in total; S5: input the three effective feature layers into the detection head to complete target detection and recognition; the detection head is composed of two branches, and the two branches are used to extract the location information and category information of the object respectively; S6: Use cross-validation to optimize model parameters and use a composite loss function to optimize the prediction accuracy of both location and category. S7: Use the test set to evaluate the performance of the SLM_YOLO model; and adjust the model's hyperparameters in the process to achieve the best performance of the model; S8: Deploy and apply the trained and verified SLM_YOLO model in actual usage scenarios to complete the automatic recognition and classification of tower base components.

2. A multi-scale tower crane component recognition method based on the YOLO model according to claim 1, characterized in that: The C2f_SCConv layer in the backbone network includes a convolution module, a Split module, two Bottleneck_SCConv layers and a splicing layer; the image in the enhanced dataset enters the C2f_SCConv layer for deep processing after feature extraction in the convolution layer. The specific processing method is as follows: S3.1: Receive the intermediate feature map output by the previous layer of the C2f_SCConv layer through the convolution module, and perform further convolution processing on the intermediate feature map; S3.2: The intermediate feature map after the convolution processing in the previous step is divided into two parts through the Split module, one part is passed to the final concatenation layer, and the other part is sent to two Bottleneck_SCConv layers for processing; each Bottleneck_SCConv layer contains a normal convolution and a self-calibration convolution; S3.3: When the first Bottleneck_SCConv layer processes the intermediate feature map, it continuously outputs the processing result through a normal convolution and a self-calibration convolution; the output processing result of the first Bottleneck_SCConv layer is divided into two parts, one part enters the next Bottleneck_SCConv layer for processing, and the other part enters the final splicing layer; the second Bottleneck_SCConv layer also continuously outputs the processing result through a normal convolution and a self-calibration convolution; the output processing result of the second Bottleneck_SCConv layer directly enters the final splicing layer.

3. A multi-scale tower crane component recognition method based on the YOLO model according to claim 2, characterized in that: The specific implementation process of the self-calibration convolution is as follows: S3.3.1: Average pooling: Let the input to the self-calibration convolution intermediate feature map be X1, the filter size be r×r, and the stride be r; then: T1=AvgPool r (X1); Where T1 is the downsampled intermediate feature map obtained by applying an average pooling function of size r×r and stride r; AvgPool r is the average pooling function, S3.3.2: Feature enhancement and upsampling: X'1=Up(F2(T1))=Up(T1*K2); Where X′1 is the intermediate feature map obtained by applying K2 operation to T1; Up() is a bilinear interpolation operator, F2(T1)=T1*K2, where K2 includes an average pooling operation and a convolution operation, and F2(T1) means applying K2 to T1; S3.3.3: Self-calibration operation: Y′1=F3(X1)·σ(X1+X′1); Where Y′1 is the self-calibration output; F3(X1)=X1*K3, where K3 represents the convolution operation, F3(X1) represents applying the convolution K3 to X1, and σ is the sigmoid activation function; S3.3.4: Final convolution output: Y1=F4(Y′1)=Y′1*K4 Among them, Y1 is the final output of the self-calibration convolution, K4 represents the final convolution operation, and F4(Y′1) represents the application of convolution K4 to Y′1.

4. The multi-scale tower crane component recognition method based on the YOLO model according to claim 1 is characterized in that: The FPN structure in the neck network includes an SPPF_LSKA layer, two upsampling layers, two concatenation layers and two C2f_SCConv layers; the PAN structure includes two convolution layers, two concatenation layers and two C2f_SCConv layers.

5. A multi-scale tower crane component recognition method based on the YOLO model according to claim 4, characterized in that: The feature map finally outputted in the backbone network enters the SPPF_LSKA layer; the SPPF_LSKA layer includes two convolutional layers, three pooling layers, an LSKA layer and a concatenation layer; the processing process of the feature map by the SPPF_LSKA layer is as follows: S4.1: The feature map enters the convolution layer in the SPPF_LSKA layer for preliminary feature extraction; part of the feature map processed by the convolution layer is sent to the splicing layer, and the other part enters the subsequent pooling layer; S4.2: The part fed into the pooling layer in the previous step enters three consecutive maximum pooling layers, each of which captures features of different scales through pooling windows of different sizes; Part of the output of the first and second pooling layers goes directly to the concatenation layer, and the other part goes to the subsequent pooling layer; the output of the third pooling layer goes to the LSKA layer; S4.3: The part entering the LSKA layer first passes through two depthwise separable convolutions and then enters two dilated convolution layers, thereby increasing the spatial coverage of the feature map; the output of the LSKA layer enters the subsequent concatenation layer; S4.4: The concatenation layer integrates and fuses the feature maps received in all the previous steps, and inputs the integrated feature maps into the final convolutional layer to weight the feature maps to generate weighted feature maps.

6. A multi-scale tower crane component recognition method based on the YOLO model according to claim 5, characterized in that: After generating the weighted feature map, the SPPF_LSKA layer sends the weighted feature map to the subsequent FPN and PAN structures of the neck network to extract three effective feature layers.

7. The multi-scale tower crane component recognition method based on the YOLO model according to claim 1 is characterized in that: In step S5, the branch for extracting the object's position information extracts the object's position information through a 3x3 convolution, a MultiSEAM layer, a 1x1 convolution and Bbox.loss; The branch used to extract object category information extracts the object's category information through a 3x3 convolution, a MultiSEAM layer, a 1x1 convolution, and Cls.loss.

Images