Coaxial line joint detection method based on improved YOLO11 model

Abstract

The invention discloses a coaxial line joint detection method based on an improved YOLO11 model. The method comprises the following steps: making a coaxial line joint target detection data set; inputting a coaxial line joint target detection data set into the improved YOLO11 model for training to obtain an optimal improved YOLO11 model with stable prediction precision as a coaxial line joint detection model; inputting to-be-detected data into the obtained coaxial line joint detection model to perform accurate detection and positioning of the coaxial line joint, screening an effective detection frame through non-maximum suppression, rejecting a low-credibility prediction result in combination with a confidence coefficient threshold, and finally outputting a detection image marked with a joint position and confidence coefficient. And positioning and identification of the coaxial line joint are completed. The method solves the problems of data set scarcity, difficulty in small target identification, strong reflection interference, limited model deployment and the like in coaxial line joint detection, gives consideration to detection precision and real-time performance, and is suitable for a 3C product automatic assembly line scene.

Description

Technical Field

[0001] This invention relates to the technical field of machine vision inspection, and in particular to a method for detecting coaxial cable connectors based on an improved YOLO11 model. It is applicable to industrial scenarios such as automated assembly lines for 3C products and can achieve high-precision real-time detection and positioning of small-sized, low-texture, and highly reflective coaxial cable connectors. Background Technology

[0002] With the rapid development of intelligent manufacturing and the 3C industry, 3C products are exhibiting significant characteristics of miniaturization, high density, and complex structure. The connection accuracy of coaxial cable connectors during assembly directly determines the overall performance and stability of the product. As a core component of electronic assembly lines, coaxial cable connectors are characterized by their small size, lack of surface texture, and susceptibility to reflection, leading to numerous technical bottlenecks in their detection and positioning.

[0003] 1. Coaxial cable connectors belong to the category of small industrial targets. There are almost no relevant samples in public datasets. Self-collection requires covering multiple devices and multiple working conditions, which is difficult to operate. Insufficient sample diversity can easily lead to model overfitting and make it difficult to adapt to the dynamic changes in industrial scenarios.

[0004] 2. In industrial settings, lighting conditions fluctuate, there are large-scale production line lighting changes, external natural light interference, and the metal material on the surface of the coaxial cable connector is prone to reflection, causing confusion between the gray values ​​of the target and the background. Traditional detection models are prone to "background misjudgment" or "target missed detection".

[0005] 3. The coaxial cable connector is too small. At a standard image resolution, such as 1280×720, the target pixel ratio is less than 0.5%. Existing models have weak feature extraction capabilities for small targets, and key details such as the edge of the connector port are easily buried by noise.

[0006] 4. Real-time inspection requirements on assembly lines place stringent demands on model inference speed. Meanwhile, industrial edge devices have limited computing power. Existing complex models have large parameter sets and time-consuming inference, while lightweight models struggle to balance detection accuracy, making adaptation difficult. Although YOLO11, as a mainstream real-time target detection framework, offers significant improvements in detection speed and accuracy compared to its predecessors, its direct application to coaxial cable joint inspection still presents challenges such as insufficient focus on small target features, blurred edge features, and difficulty in balancing accuracy and lightweight design. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings and deficiencies of the prior art and propose a coaxial cable joint detection method based on an improved YOLO11 model. This method can effectively address the problems of scarce datasets, difficulty in identifying small targets, blurred edge features, and limited model deployment in coaxial cable joint detection, and meet the real-time requirements of industrial assembly lines while ensuring detection accuracy.

[0008] To achieve the above objectives, the technical solution provided by this invention is as follows: a coaxial cable joint detection method based on an improved YOLO11 model. The improved YOLO11 model modifies the backbone network, feature fusion network, and model structure of the original YOLO11 model to achieve accurate detection and localization of coaxial cable joints. Specifically, the improvement to the backbone network involves replacing the C2PSA downsampling module with a highly efficient channel attention module (EA) that possesses dynamic channel adaptation and small target feature enhancement capabilities, thereby enhancing the focusing ability of small target features. The improvement to the feature fusion network involves embedding a lightweight edge enhancement module (LEE) based on a depthwise separable convolution module in the upsampling stage of the feature fusion network, thereby enhancing edge feature response. The improvement to the model structure involves employing a layer-by-layer channel optimization strategy based on the importance of each layer, calculating the importance value of each layer, and adjusting the number of channels in each layer according to this importance value, thus achieving model lightweighting while maintaining accuracy.

[0009] The specific implementation of the coaxial cable joint detection method includes the following steps:

[0010] 1) Collect images of coaxial cable connectors from multiple devices and in multiple scenarios and perform preprocessing to obtain a coaxial cable connector target detection dataset with sufficient sample diversity;

[0011] 2) The coaxial cable joint target detection dataset was input into the improved YOLO11 model for training. The optimal improved YOLO11 model with stable prediction accuracy was obtained as the coaxial cable joint detection model. During training, multi-scale features of coaxial cable joints were extracted by embedding an EA backbone network. The input feature map was first compressed and aggregated with channel information, then the dynamic receptive field module was fused to adjust the feature capture range, and finally attention weights were generated to strengthen key features. The feature fusion network with embedded LEE was used to optimize feature representation to extract edge information. The feature integrity was preserved by combining residual connections and activation functions, and then the channel dimension was matched by point convolution. At the same time, the model was lightweighted based on layer importance. The values ​​of three factors, including accuracy, number of parameters and latency, were quantitatively evaluated for each layer, i.e., the "accuracy-number of parameters-latency" ternary objective. The pruning rate was assigned according to importance to achieve channel pruning.

[0012] 3) Input the data to be detected into the coaxial cable joint detection model obtained in step 2) to accurately detect and locate the coaxial cable joint. Then, use non-maximum suppression to filter valid detection boxes and combine confidence threshold to remove low confidence prediction results. Finally, output the detection image with the joint position and confidence level marked to complete the location and identification of the coaxial cable joint.

[0013] Furthermore, step 1) includes the following steps:

[0014] Multi-device data acquisition: The core device is a D435i depth camera. During acquisition, the camera height is fixed, and the shooting angle covers the range of 0° to 45°, simultaneously acquiring RGB images and depth information; auxiliary devices include smartphones and web crawling; the acquisition scenarios cover 3 typical backgrounds and 5 light intensities of industrial assembly lines.

[0015] Image preprocessing: Remove severely blurred, overexposed or underexposed images, and images without a target; then remove image noise by using a Gaussian filter with a convolution kernel size of 3×3, adjust image contrast by adaptive histogram equalization, and eliminate the influence of lighting differences by brightness normalization to ensure that the preprocessed image is clear and has a uniform grayscale distribution.

[0016] High-precision image annotation: Use the LabelImg annotation tool to annotate the bounding box of the coaxial cable connector in YOLO format. The annotation rule is: the bounding box must completely enclose the connector area.

[0017] Data augmentation: Various augmentation strategies are implemented on the labeled images to improve sample diversity and model generalization ability, ultimately resulting in a coaxial connector target detection dataset with sufficient sample diversity.

[0018] Furthermore, the backbone network embedded with the EA performs the following operations:

[0019] First, regarding the input dimension as The feature maps are subjected to global average pooling, where , For the feature map height and width, The total number of channels is determined by analyzing all spatial positions within each channel. The pixel values ​​are averaged. These represent the pixel values ​​on the x and y axes of the image, respectively, which were originally distributed across... Spatial dimension feature information is compressed and aggregated into A one-dimensional channel descriptor is used to effectively extract the global feature information of each channel and discard redundant details in the spatial dimension. Then, to avoid the surge in computational complexity and feature overfitting caused by an excessive number of channels, a 1×1 convolution operation is used to compress the number of channels in the one-dimensional channel descriptor to [a smaller number]. , / / represents integer division, To reduce the number of channels after dimensionality reduction, this channel compression design has dual technical advantages: firstly, it achieves reasonable simplification of the number of channels through integer division, reducing subsequent computational pressure; secondly, it forces the number of intermediate layer channels to be no less than 8 through the max function, ensuring that key feature information is not lost due to excessive dimensionality reduction during channel compression, thus guaranteeing the integrity of feature representation. Subsequently, a multi-level progressive activation mechanism is constructed, consisting of an unbiased linear layer, a memory-optimized ReLU activation function, and a sigmoid activation function. The design of the unbiased linear layer avoids the interference of bias parameters on channel dependency learning. The memory-optimized ReLU activation function maintains non-linear feature extraction capabilities while achieving in-situ calculation of feature maps through the inplace=True parameter setting, significantly reducing memory usage. Finally, the sigmoid activation function maps the output of the multi-level progressive activation mechanism to the [0,1] interval, generating channel attention weights that accurately reflect the importance of each channel. Finally, the generated channel attention weights are compared with the original input... The feature map is subjected to pixel-level multiplication operations on a channel-by-channel basis to weight and enhance the channels containing key features of the coaxial cable connector, while suppressing redundant information such as background noise and irrelevant textures, thereby improving the distinguishability between the target and the background in the feature map.

[0020] The EA introduces a delayed fully connected layer initialization mechanism, which sets the fully connected layer parameter self.fc to None during the module initialization phase. Only when a specific input feature map is received during the first forward propagation does the adapted fully connected layer structure dynamically create based on the actual number of channels in the input feature map. This allows it to be integrated into variant networks of the YOLO11 model with different parameter scales such as n, s, m, l, and x, without needing to adjust the module structure for different network variants.

[0021] The mathematical formulas corresponding to the entire operation process described above are as follows:

[0022] ;

[0023] ;

[0024] ;

[0025] ;

[0026] ;

[0027] In the formula, For the input feature map at location Number of channels pixel values, Number of channels The global average pooling result, This represents the feature vector after dimensionality reduction by an unbiased linear layer following the global pooling result. This indicates a linear layer without bias. Represents bias. Represents non-biased. For bias, This represents the result of global average pooling. This indicates that the number of output channels of the linear layer after dimensionality reduction is a fixed value. , The number of output channels of a linear layer is equal to the total number of channels of the input feature map. , This represents the memory-optimized ReLU activation function. The attention weights of the channels, For the first Attention weights for each channel, For in position Number of channels The output feature map after weight enhancement;

[0028] Through the aforementioned backbone network, four downsampling operations are performed sequentially during feature extraction, with downsampling rates of 2, 4, 8, and 16, respectively. By gradually increasing the channel dimension of the feature map and reducing the spatial resolution, a progressive extraction from shallow texture features to deep semantic features is achieved, ultimately resulting in four levels of output feature maps from shallow to deep, denoted as P3, P4, P5, and P6, respectively. Among them, P3 has the highest spatial resolution, which can accurately capture the subtle textures and edge details of the coaxial cable connector, while P4 enhances the semantic feature expression while preserving the spatial resolution. The two work together to optimize the feature extraction needs of small targets such as coaxial cable connectors.

[0029] Furthermore, the feature fusion network embedded with LEE retains the original structural framework of the original YOLO11 model to ensure the compatibility of the network structure and basic detection performance.

[0030] The core workflow of LEE consists of three closely linked technical steps, which work together to achieve efficient extraction and optimization of edge features, as detailed below:

[0031] The first step is efficient edge feature extraction, which adopts an innovative design using a depthwise separable convolutional architecture. This completely eliminates the redundant computations in the feature extraction process of traditional standard convolutions. First, the input feature map is processed through a 3×3 depthwise convolutional layer. The convolution kernel of this 3×3 depthwise convolutional layer corresponds one-to-one with the number of channels in the feature map. That is, each convolution kernel performs spatial dimension convolution operations only within a single channel, which can accurately focus on the spatial position information and texture features of the coaxial cable joint edge. After completing the single-channel spatial edge feature extraction, cross-channel feature fusion is performed through a 1×1 point convolutional layer. This 1×1 point convolutional layer can integrate the edge features extracted from each channel individually, and flexibly adjust the channel dimension of the feature map according to the needs of subsequent feature processing, so as to achieve cross-channel complementarity and dimension adaptation of edge features.

[0032] The second stage is dynamic adaptive channel adjustment, which innovatively introduces a dual-delay initialization mechanism. That is, the 1×1 point convolutional layer and the 3×3 depth convolutional layer in the module do not predefine the specific number of input and output channels and convolutional kernel parameters during the module initialization stage. Instead, when the module performs its first forward propagation, it dynamically creates an adapted convolutional layer structure and convolutional kernel parameters based on the channel dimension and spatial size of the actual received input feature map. This completely breaks the fixed limitation of the number of input feature map channels in the traditional edge enhancement module, enabling it to flexibly adapt to the multi-scale model of YOLO11 model with different parameter scales such as n, s, m, l, and x without any structural modification.

[0033] The third step is feature integration and performance optimization. It uses dual technical means to ensure the integrity of features and the efficiency of model operation during the edge enhancement process. On the one hand, a residual connection mechanism is introduced to directly add and fuse the edge features extracted by depthwise separable convolution with the original input features of the module. This can effectively avoid the loss of basic feature information during the edge feature enhancement process and ensure that the feature map contains both accurate edge details and retains the complete overall features of the target. On the other hand, the SiLU activation function is used to replace the traditional ReLU activation function. The SiLU activation function has smooth nonlinear characteristics, which can enhance the nonlinear expression of features while optimizing the gradient flow of the model.

[0034] Furthermore, the LEE performs the following operations:

[0035] For feature maps of arbitrary dimensions input to the feature fusion network First, spatial edge features are extracted using a dynamically initialized 3×3 depth convolutional layer. This 3×3 depth convolutional layer is based on... The actual number of channels dynamically generates a corresponding number of 3×3 convolutional kernels, performing convolution operations only within a single channel to accurately capture key edge features such as the port contour and metal seam of the coaxial cable connector, while avoiding redundant cross-channel computations in traditional convolution. Next, a dynamically initialized 1×1 point convolutional layer is used to perform cross-channel fusion of the edge feature maps extracted by depth convolution, integrating edge information from different channels and initially adjusting the channel dimensions of the feature maps. Finally, a residual connection mechanism is activated to combine the edge features processed by point convolution with the original input features of the module. Pixel-by-pixel addition is performed to obtain the feature map after residual fusion. Ensure that basic features are not lost; then... The input is fed into the SiLU activation function, whose non-linear transformation enhances the discriminative power of the features, further strengthening the difference between the coaxial cable connector edge features and the background. Finally, another dynamically initialized 1×1 point convolutional layer is used to fine-tune the number of channels in the activated feature map, ensuring the final output feature map... The input dimensions of the feature fusion network are fully matched with those of the subsequent feature fusion layers, ensuring the smoothness of the entire feature fusion network in the feature processing flow.

[0036] The mathematical formulas corresponding to the entire operation process described above are as follows:

[0037] ;

[0038] ;

[0039] ;

[0040] ;

[0041] ;

[0042] In the formula, A 3×3 depthwise convolution kernel dynamically initialized for each channel. Where is the row index of the convolution kernel. The column index of the convolution kernel. The input channel index bound to the convolution kernel; To extract the spatial location x-coordinate of depthwise convolution. The vertical axis is The number of channels is Edge feature map; The spatial location of the input feature map; This is the output feature map of the first 1×1 pointwise convolution; , These are dynamically initialized 1×1 point convolution weight matrices, the former used to fuse channel information and the latter used to match the channel dimensions of subsequent layers. This is the edge enhancement feature map after SiLU activation; Use the Sigmoid activation function; This is the feature map after residual fusion. This is the final output edge enhancement feature map.

[0043] Furthermore, the layer-by-layer channel optimization strategy does not employ traditional uniform or random pruning methods. Instead, it intelligently allocates the pruning rate for each layer through hierarchical importance quantification calculation, and the pruning range is constrained. This process achieves precise reduction of model parameters and computational load while maximizing the preservation of model detection accuracy. The implementation steps of this layer-by-layer channel optimization strategy are as follows:

[0044] The first step is to perform hierarchical metric calculations. First, we clarify three optimization objectives: accuracy, number of parameters, and latency. This covers the core evaluation dimensions of model performance and avoids performance imbalances caused by optimizing a single dimension. Specifically, we use the value mAP@0.5 from the dataset. This value represents the average accuracy when the intersection-over-union (IoU) threshold is 0.5, serving as a core proxy metric for model accuracy. mAP@0.5 directly reflects the model's precision and recall capabilities for detecting coaxial cable joints and is a recognized key accuracy evaluation standard in object detection tasks. By traversing the parameter matrices of each layer of the model, we count the number of parameters in each layer, including learnable parameters such as convolution kernel parameters and bias parameters, as a core metric for measuring the computational complexity of that layer. Finally, we conduct actual inference tests on the NVIDIA Jetson AGXXavier edge computing platform, recording the individual inference time of each network layer using a high-precision timer as a core metric for measuring the real-time performance of that layer.

[0045] After acquiring the ternary target data for each layer, normalization is performed on each type of indicator to eliminate the influence of differences in dimensions between different indicators. The specific normalization method is as follows: for precision indicators... , representing the first The accuracy contribution value corresponding to the layer was obtained by testing using the layer-by-layer masking method, and then mapped to the [0,1] interval using min-max normalization; for parameter indicators, the first... Layer parameter quantity Similarly, min-max normalization is used to map to the [0,1] interval; for the latency metric, the first... Layer inference time The reverse min-max normalization is used to map to the [0,1] interval. The purpose of reverse normalization is to make the layer with shorter latency get a higher normalization score, which is consistent with the optimization direction of accuracy and parameter quantity indicators.

[0046] Subsequently, to balance the importance of the ternary objective data, weighting coefficients α, β, and γ are introduced, where α, β, and γ are preset non-negative weight values ​​based on actual needs, satisfying α + β + γ = 1. The normalized ternary objective data are then linearly weighted and summed to obtain the hierarchical importance of each level. Its calculation formula is The design logic of this formula is: the layer that contributes more to accuracy, The larger the value, the fewer the number of parameters in the layer. The larger the value, The larger the value, the shorter the inference time of the network layer. The larger the value, The larger the value, the more accurately the overall importance of each layer can be quantified through weighted summation.

[0047] The above-mentioned importance of the entire hierarchy The mathematical formula for quantitative calculation is as follows:

[0048] ;

[0049] In the formula, , , These represent the maximum values ​​of each indicator across all layers. Contribution to accuracy For parameters, For inference delay;

[0050] The second step is intelligent allocation of differentiated pruning rates, after obtaining the hierarchical importance of all layers. Then, apply the following to all layers: The values ​​are globally sorted from largest to smallest to clarify the importance level of each layer in the entire model. Based on the sorting results, a differentiated pruning strategy is implemented: for shallow networks with low layer importance, which are responsible for extracting basic texture features and have many redundant parameters, but have little impact on the final detection accuracy, a high pruning rate is assigned, that is, their number of channels is significantly reduced to minimize the computational cost of the model; for deep networks with high layer importance, which are responsible for extracting high-level semantic features and play a decisive role in the target detection accuracy, they have few redundant parameters, a low pruning rate is assigned, and only a small number of redundant channels are reduced to retain their feature extraction capabilities to the maximum extent.

[0051] The third step is to constrain the pruning range. To avoid a severe decrease in feature extraction capability due to an excessively high pruning rate, or an insignificant lightweighting effect due to an excessively low pruning rate, a cutoff function is introduced. Limit the pruning rate of each layer to a range and set a minimum pruning rate. With the maximum value The final calculation of the pruning rate is done using the formula. Implementation, in which For the final pruning rate, This is the pruning intensity adjustment coefficient, used to control the overall aggressiveness of pruning. This represents the pruning baseline coefficient, which is inversely proportional to the importance of the hierarchy. That is, the lower the importance, the larger the pruning baseline coefficient, and the higher the pruning rate. Then, through... The function limits the calculated pruning rate to [ , Within the specified range, ensure that the pruning operation can effectively reduce the model complexity without significantly affecting the detection accuracy.

[0052] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0053] 1. The coaxial cable joint detection dataset constructed in this invention adopts multi-device heterogeneous acquisition and multi-level data augmentation, covering 3 typical backgrounds and 5 light intensities, with sufficient sample diversity; the annotation accuracy is controlled within 1 pixel and has undergone double verification, resulting in high data quality. It effectively solves the problems of scarce public datasets and poor sample fit, providing reliable data support for model training.

[0054] 2. This invention addresses the difficulty of feature extraction caused by the small size and low pixel ratio of coaxial cable connectors. The designed EA focuses on the key features of small targets by global information aggregation and precise channel weight allocation, suppressing background noise interference. This solves the problem of weak response of traditional models to industrial small target features and improves the reliability of small target detection.

[0055] 3. This invention addresses the problem of target and background confusion caused by large fluctuations in lighting in industrial scenarios and the easy reflection of metal surfaces of coaxial cable connectors. The designed LEE enhances edge feature extraction through depthwise separable convolution, preserves feature integrity by combining residual connections, and adapts to multi-scale models with adaptive channel adjustment, effectively distinguishing reflective areas from the target body and reducing the risk of false detection and false negative detection in complex environments.

[0056] 4. This invention addresses the contradiction between the limited computing power of industrial edge devices and the real-time inspection requirements of assembly lines. Through the aggressive dimensionality reduction design of EA, the efficient convolutional architecture of LEE, and the layer importance-driven pruning strategy, the number of model parameters and computational load are reduced without significantly sacrificing detection accuracy. This solves the adaptation problem of slow inference of complex models and low accuracy of lightweight models, and meets the real-time deployment requirements of industrial sites.

[0057] 5. The EA and LEE proposed in this invention both have dynamic adaptation characteristics and can be migrated to other industrial small target detection scenarios without significant adjustments, solving the problem of weak generalization ability of dedicated detection models; at the same time, the model can be directly deployed to industrial edge devices and seamlessly integrated with robot assembly guidance systems, providing technical support for the automated assembly of 3C products and similar electronic devices, and reducing the cost of manual intervention.

[0058] 6. The average accuracy (mAP@0.5) of this invention reaches 95.3%, which is higher than the original YOLO11, with better generalization ability and stronger detection capability for small targets such as coaxial cable connectors. Meanwhile, the improved YOLO11 model has only 2.03M parameters, achieving 16.7 FPS on NVIDIA Jetson AGX Xavier industrial edge devices. After TensorRT quantization, it can run stably on industrial computing platforms with lower computing power, fully meeting the real-time detection requirements of coaxial cable connectors in 3C product assembly lines, providing accurate and reliable technical support for automated assembly. Attached Figure Description

[0059] Figure 1 shows the network structure of the improved YOLO11 model. In the figure, the input is the input image, Conv is the basic convolutional module, C3K2 is the residual convolutional module, C2PSA is the original downsampling module of the YOLO11 model, EfficientAttention is the efficient channel attention module that replaces C2PSA, abbreviated as EA, SPPF is the multi-scale enhancement module, Backbone is the backbone network, Neck is the feature fusion network (including upsampling and fusion operation Concat), LightEdgeEnhanced is the lightweight edge enhancement module, abbreviated as LEE, which is embedded in the Neck upsampling stage, Head is the detection head (including the detection module Detect), and the output is the model output result, including the predicted bounding box coordinates and class probability.

[0060] Figure 2 is a schematic diagram of the efficient channel attention module (EA). In the figure, the input feature dimension is B×C×H×W, where B is the batch size, C is the number of channels, H is the height, W is the width, pool is the global average pooling operation, 1×1 Conv is an unbiased 1×1 convolutional layer, FC is an unbiased fully connected layer, ReLU is a memory-optimized ReLU activation function, Sigmoid is the activation function, the weights are multiplied channel by channel with the original input feature map, and the output is the enhanced feature map.

[0061] Figure 3 is a schematic diagram of the lightweight edge enhancement module (LEE); in the figure, 1×1 Conv1 is a 1×1 point convolutional layer, SiLU is the SiLU activation function to enhance nonlinear expression, and the output is an enhanced feature map.

[0062] Figure 4 shows the experimental results of an example of the present invention; in the figure, label "1" represents the coaxial connector of the object being measured, and it can be correctly identified even when placed in different orientations on different planes. Detailed Implementation

[0063] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0064] As shown in Figures 1 to 3, this embodiment discloses a coaxial cable joint detection method based on an improved YOLO11 model. This improved YOLO11 model modifies the backbone network, feature fusion network, and model structure of the original YOLO11 model to achieve accurate detection and localization of coaxial cable joints. Specifically, the backbone network is improved by replacing the downsampling module C2PSA with a high-efficiency channel attention module (EA) that possesses dynamic channel adaptation and small target feature enhancement capabilities, thereby enhancing the focusing ability of small target features. The feature fusion network is improved by embedding a lightweight edge enhancement module (LEE) based on a depthwise separable convolution module in the upsampling stage of the feature fusion network, thereby enhancing edge feature response. The model structure is improved by employing a layer-by-layer channel optimization strategy that differentiates by the importance of each layer. The importance value of each layer is calculated separately, and the number of channels in each layer is adjusted according to this importance value, achieving model lightweighting while maintaining accuracy.

[0065] The specific implementation of this coaxial cable joint testing method includes the following steps:

[0066] 1) Acquire and preprocess coaxial cable connector images from multiple devices and scenarios to obtain a coaxial cable connector target detection dataset with sufficient sample diversity, including the following steps:

[0067] 1.1) Multi-device data acquisition: A heterogeneous acquisition strategy of "core device + auxiliary device" was adopted. The core device was a D435i depth camera (1280×720 resolution, 30FPS frame rate). During acquisition, the camera height was fixed at 50cm and the shooting angle was 0°-45°, and RGB images and depth information were acquired simultaneously. The auxiliary devices included two smartphones of different brands (to simulate low and medium image quality respectively) and web crawling (to supplement samples of special working conditions such as strong reflection and occlusion). The acquisition scenes covered three typical backgrounds of industrial assembly lines (metal tabletop, plastic tray, circuit board) and five light intensities (50lux, 200lux, 500lux, 800lux, 1000lux), and a total of 1,500 original images were acquired.

[0068] 1.2) Image preprocessing: Low-quality images (severely blurred, overexposed / underexposed, without a target) are removed, and a total of 1000 valid images are retained; then, Gaussian filtering (3×3 kernel size) is used to remove image noise, adaptive histogram equalization is used to adjust image contrast, and brightness normalization is used to eliminate the influence of illumination differences, so as to ensure that the preprocessed images are clear and have uniform grayscale distribution.

[0069] 1.3) High-precision image annotation: Use the LabelImg annotation tool to annotate the bounding box of the coaxial cable connector in YOLO format. The annotation rule is: the bounding box must completely enclose the connector area.

[0070] 1.4) Multi-level data augmentation: A combined strategy of "basic augmentation + advanced augmentation" is implemented for the labeled images to improve sample diversity and model generalization ability, ultimately obtaining a coaxial connector target detection dataset with sufficient sample diversity. The data augmentation methods include: image scaling, image flipping, random brightness enhancement, and Mosaic data augmentation. Mosaic data augmentation is an image data augmentation technique that increases the diversity of training data and improves the model's generalization and robustness to different scales and backgrounds by stitching multiple images into one image.

[0071] 2) The coaxial cable joint target detection dataset was input into the improved YOLO11 model for training. The optimal improved YOLO11 model with stable prediction accuracy was obtained as the coaxial cable joint detection model. During training, multi-scale features of coaxial cable joints were extracted by embedding an EA backbone network. The input feature map was first compressed and aggregated with channel information, then the dynamic receptive field module was fused to adjust the feature capture range, and finally attention weights were generated to strengthen key features. The feature fusion network with embedded LEE was used to optimize feature representation to extract edge information. The feature integrity was preserved by combining residual connections and activation functions, and then the channel dimension was matched by point convolution. At the same time, the model was lightweighted based on layer importance. The numerical values ​​of each layer, including accuracy, number of parameters, and latency, were quantitatively evaluated, i.e., the "accuracy-number of parameters-latency" ternary objective. The pruning rate was assigned according to importance to achieve channel pruning.

[0072] Images (640×640) from the coaxial cable connector target detection dataset are input into the backbone network embedded with EA. The core improvement of the backbone network is to replace the original C2PSA with EA. This EA enhances the ability to capture features of small targets through a coherent process of "feature compression - dynamic receptive field fusion - attention weight generation - feature weighting".

[0073] First, regarding the input dimension as The feature maps are subjected to global average pooling, where , For the feature map height and width, The total number of channels is determined by analyzing all spatial positions within each channel. The pixel values ​​are averaged. These represent the pixel values ​​on the x and y axes of the image, respectively, which were originally distributed across... Spatial dimension feature information is compressed and aggregated into A one-dimensional channel descriptor is used to effectively extract the global feature information of each channel and discard redundant details in the spatial dimension. Then, to avoid the surge in computational complexity and feature overfitting caused by an excessive number of channels, a 1×1 convolution operation is used to compress the number of channels in the one-dimensional channel descriptor to [a smaller number]. , / / represents integer division, To reduce the number of channels after dimensionality reduction, this channel compression design has dual technical advantages: firstly, it achieves reasonable simplification of the number of channels through integer division, reducing subsequent computational pressure; secondly, it forces the number of intermediate layer channels to be no less than 8 through the max function, ensuring that key feature information is not lost due to excessive dimensionality reduction during channel compression, thus guaranteeing the integrity of feature representation. Subsequently, a multi-level progressive activation mechanism is constructed, consisting of an unbiased linear layer, a memory-optimized ReLU activation function, and a sigmoid activation function. The design of the unbiased linear layer avoids the interference of bias parameters on channel dependency learning. The memory-optimized ReLU activation function maintains non-linear feature extraction capabilities while achieving in-situ calculation of feature maps through the inplace=True parameter setting, significantly reducing memory usage. Finally, the sigmoid activation function maps the output of the multi-level progressive activation mechanism to the [0,1] interval, generating channel attention weights that accurately reflect the importance of each channel. Finally, the generated channel attention weights are compared with the original input... The feature map is subjected to pixel-level multiplication operations on a channel-by-channel basis to achieve weighted enhancement of channels containing key features of the coaxial cable connector, while suppressing redundant information such as background noise and irrelevant textures, thereby significantly improving the distinguishability between the target and the background in the feature map.

[0074] The EA innovatively introduces a delayed fully connected layer initialization mechanism. This involves setting the fully connected layer parameter `self.fc` to `None` during the module initialization phase. Only when a specific input feature map is received during the first forward propagation does the EA dynamically create an adapted fully connected layer structure based on the actual number of channels in the input feature map. This design overcomes the technical limitation of traditional attention modules requiring pre-specified input channel numbers, enabling seamless integration into variant networks of the YOLO11 model with different parameter scales (n, s, m, l, x). It eliminates the need for module structure adjustments for different network variants, significantly improving the module's engineering practicality, structural flexibility, and environmental robustness. The corresponding formula is as follows:

[0075] ;

[0076] ;

[0077] ;

[0078] ;

[0079] ;

[0080] In the formula, For the input feature map at location Number of channels pixel values, Number of channels The global average pooling result, This represents the feature vector after dimensionality reduction by an unbiased linear layer following the global pooling result. This indicates a linear layer without bias. Represents bias. Represents non-biased. For bias, This represents the result of global average pooling. This indicates that the number of output channels of the linear layer after dimensionality reduction is a fixed value. , The number of output channels of a linear layer is equal to the total number of channels of the input feature map. , This represents the memory-optimized ReLU activation function. The attention weights of the channels, For the first Attention weights for each channel, For in position Number of channels The output feature map after weight enhancement;

[0081] Through the aforementioned backbone network, four downsampling operations are sequentially performed during feature extraction, with downsampling rates of 2, 4, 8, and 16, respectively. By progressively increasing the channel dimension of the feature map and reducing the spatial resolution, a progressive extraction from shallow texture features to deep semantic features is achieved, ultimately resulting in four levels of output feature maps from shallow to deep, denoted as P3 (size 80×80×256), P4 (size 40×40×512), P5 (size 20×20×1024), and P6 (size 10×10×2048), corresponding to the feature information of small, medium, large, and ultra-large scale targets, respectively. Among them, P3 has the highest spatial resolution, which can accurately capture the subtle texture and edge details of the coaxial cable connector. P4 enhances the semantic feature expression while retaining a certain spatial resolution. The two work together to optimize the feature extraction needs of small targets such as coaxial cable connectors, completely solving the technical defect of insufficient feature capture in the original backbone network during small target detection.

[0082] The feature fusion network with embedded LEE adopts the original YOLO11 structure framework, but embeds a lightweight edge enhancement module (LEE) in the upsampling stage. The LEE is designed with "efficient edge extraction - adaptive channel adaptation - parameter optimization" as its core design idea. It enhances the edge feature discrimination of coaxial cable connectors (such as MHF4 connectors) under the premise of lightweight design, and has cross-model adaptation capability, which perfectly meets the needs of real-time detection.

[0083] The core workflow of this LEE consists of three parts. First, efficient edge feature extraction is achieved through a depthwise separable convolutional architecture—3×3 depthwise convolution performs spatial convolution within a single channel, accurately capturing the spatial features of the connector edge, reducing the number of parameters by approximately 90% compared to standard 3×3 convolution. Subsequently, 1×1 pointwise convolution completes cross-channel feature fusion, integrating edge information while adjusting feature dimensions. Second, adaptive channel adjustment is achieved through a delayed initialization mechanism (self.conv1 and self.edge_conv are dynamically created based on the input during forward propagation), eliminating the need for predefined input channel numbers. This allows for flexible adaptation to multiple scales of YOLOv11 models, such as n, s, m, l, and x, greatly improving the module's versatility. Finally, feature integration and optimization are implemented. Residual connections are introduced to fuse edge features with the original input features, ensuring information integrity. The SiLU activation function optimizes gradient flow, enabling the module to operate efficiently on resource-constrained industrial edge devices.

[0084] In summary, the core workflow of this LEE can be divided into three closely linked technical stages, which work together to achieve efficient extraction and optimization of edge features, as detailed below:

[0085] The first stage is efficient edge feature extraction, which employs an innovative design using a depthwise separable convolutional architecture, completely eliminating redundant computations in the feature extraction process of traditional standard convolutions. Specifically, the input feature map is first processed through a 3×3 depthwise convolutional layer. The convolutional kernels of this depthwise convolutional layer correspond one-to-one with the number of channels in the feature map, meaning that each kernel performs spatial convolution operations only within a single channel. This allows for precise focusing on the spatial location information and texture features of the coaxial cable connector edge. Compared to the traditional standard 3×3 convolution, which requires simultaneous cross-channel and spatial convolution, this module reduces the number of parameters by approximately 90% while maintaining the accuracy of edge feature extraction, significantly improving computational efficiency. After completing the single-channel spatial edge feature extraction, cross-channel feature fusion is performed through a 1×1 point convolutional layer. This point convolutional layer integrates the edge features extracted individually from each channel and flexibly adjusts the channel dimension of the feature map according to subsequent feature processing needs, achieving cross-channel complementarity and dimensional adaptation of edge features.

[0086] The second stage is dynamic adaptive channel adjustment, which innovatively introduces a dual-delay initialization mechanism. That is, the 1×1 point convolutional layer and the 3×3 depth convolutional layer in the module do not predefine the specific number of input and output channels and convolutional kernel parameters during the module initialization stage. Instead, when the module performs its first forward propagation, the appropriate convolutional layer structure and convolutional kernel parameters are dynamically created based on the channel dimension and spatial size of the actual received input feature map. This design completely breaks the fixed limitation of the number of input feature map channels in traditional edge enhancement modules, enabling it to flexibly adapt to multi-scale models of YOLO11 with different parameter scales such as n, s, m, l, and x without any structural modifications. This greatly improves the module's versatility and engineering application scope, and avoids the tedious work of designing edge enhancement modules separately for different models.

[0087] The third stage involves feature integration and performance optimization, employing a dual-technical approach to ensure feature integrity and model efficiency during edge enhancement. Firstly, a residual connection mechanism is introduced, directly adding and fusing the edge features extracted through depthwise separable convolution with the module's original input features. This effectively avoids the loss of basic feature information during edge enhancement, ensuring the feature map contains both accurate edge details and retains complete overall target features. Secondly, the SiLU activation function replaces the traditional ReLU activation function. The SiLU activation function possesses smooth non-linear characteristics, enhancing the non-linear expressive power of features while optimizing gradient flow, avoiding gradient vanishing or exploding problems. This enables the module to operate efficiently and stably on resource-constrained industrial edge devices.

[0088] The specific implementation process is as follows: For the feature map of the input detection head... First, spatial edge features are extracted using dynamically initialized 3×3 depthwise convolutions to avoid redundant computations in traditional convolutions. Then, channel information is fused using 1×1 pointwise convolutions, while residual connections are introduced to add the edge features to the original input features, preventing the loss of basic features during edge enhancement. Next, the SiLU activation function is used to enhance the non-linear expressive power of the features, strengthening the feature response of joint edges (such as port contours and metal seams). Finally, dynamically initialized 1×1 convolutions are used to fine-tune the number of channels, ensuring that the output features perfectly match the dimensions of subsequent fusion layers in the detection head. The corresponding formulas for the above process are as follows:

[0089] ;

[0090] ;

[0091] ;

[0092] ;

[0093] ;

[0094] In the formula, A 3×3 depthwise convolution kernel dynamically initialized for each channel. Where is the row index of the convolution kernel. The column index of the convolution kernel. The input channel index bound to the convolution kernel; To extract the spatial location x-coordinate of depthwise convolution. The vertical axis is The number of channels is Edge feature map; The spatial location of the input feature map; This is the output feature map of the first 1×1 pointwise convolution; , These are dynamically initialized 1×1 point convolution weight matrices, the former used to fuse channel information and the latter used to match the channel dimensions of subsequent layers. This is the edge enhancement feature map after SiLU activation; Use the Sigmoid activation function; This is the feature map after residual fusion. This is the final output edge enhancement feature map.

[0095] The entire model adopts a layer-by-layer channel optimization strategy driven by layer importance. This strategy does not use the traditional uniform or random pruning method, but intelligently allocates the pruning rate of each layer through the quantitative calculation of layer importance and the process of pruning range constraint. Under the premise of maximizing the preservation of model detection accuracy, it achieves precise reduction of model parameters and computational load, and solves the technical contradiction of excessive accuracy loss or insignificant lightweighting effect in the traditional model lightweighting process.

[0096] The implementation steps of this layer-by-layer channel optimization strategy are as follows:

[0097] The first step is to perform hierarchical metric calculations. First, we clarify three optimization objectives: accuracy, number of parameters, and latency. This covers the core evaluation dimensions of model performance, avoiding performance imbalances caused by optimizing a single dimension. Specifically, we use the value mAP@0.5 from the dataset. This value represents the average accuracy when the intersection-over-union (IoU) threshold is 0.5, serving as a core proxy metric for model accuracy. mAP@0.5 directly reflects the model's precision and recall capabilities for detecting coaxial cable joints and is a recognized key accuracy evaluation standard in object detection tasks. By traversing the parameter matrices of each layer of the model, we count the number of parameters in each layer, including convolutional kernel parameters, bias parameters, and all learnable parameters, serving as a core metric for measuring the computational complexity of that layer. Finally, we conduct actual inference tests on the NVIDIA Jetson AGXXavier edge computing platform, recording the individual inference time of each network layer using a high-precision timer, serving as a core metric for measuring the real-time performance of that layer.

[0098] After acquiring the ternary target data for each layer, normalization is performed on each type of indicator to eliminate the influence of differences in dimensions between different indicators. The specific normalization method is as follows: for precision indicators... , representing the first The accuracy contribution value corresponding to the layer was obtained by testing using the layer-by-layer masking method, and then mapped to the [0,1] interval using min-max normalization; for parameter indicators, the first... Layer parameter quantity Similarly, min-max normalization is used to map to the [0,1] interval; for the latency metric, the first... Layer inference time The reverse min-max normalization is used to map to the [0,1] interval. The purpose of reverse normalization is to make the layer with shorter latency get a higher normalization score, which is consistent with the optimization direction of accuracy and parameter quantity indicators.

[0099] Subsequently, to balance the importance of the three target data, weighting coefficients α, β, and γ are introduced, where α, β, and γ are preset non-negative weight values ​​based on actual needs, satisfying α + β + γ = 1. The normalized three target data are then linearly weighted and summed to obtain the hierarchical importance of each level. Its calculation formula is The design logic of this formula is: the layer that contributes more to accuracy, The larger the value, the fewer the number of parameters in the layer. The larger the value, The larger the value, the shorter the inference time of the network layer. The larger the value, The larger the value, the more accurately the overall importance of each layer can be quantified through weighted summation.

[0100] The second step is intelligent allocation of differentiated pruning rates, after obtaining the hierarchical importance of all layers. Then, apply the following to all layers: The values ​​are globally sorted from largest to smallest to clarify the importance level of each layer in the entire model. Based on the sorting results, a differentiated pruning strategy is implemented: For shallow networks with low layer importance, these networks are mainly responsible for extracting basic texture features and have more redundant parameters, which have little impact on the final detection accuracy. A higher pruning rate is assigned, that is, their number of channels is significantly reduced to minimize the computational cost of the model. For deep networks with high layer importance, these networks are mainly responsible for extracting high-level semantic features and play a decisive role in the target detection accuracy. They have fewer redundant parameters and are assigned a lower pruning rate, reducing only a small number of redundant channels to retain their feature extraction capabilities to the maximum extent.

[0101] The third step is to constrain the pruning range. To avoid a severe decrease in feature extraction capability due to an excessively high pruning rate, or an insignificant lightweighting effect due to an excessively low pruning rate, a cutoff function is introduced. Limit the pruning rate of each layer to a range and set a minimum pruning rate. With the maximum value The final calculation of the pruning rate is done using the formula. Implementation, in which For the final pruning rate, This is the pruning intensity adjustment coefficient, used to control the overall aggressiveness of pruning. This represents the pruning baseline coefficient, which is inversely proportional to the importance of the hierarchy. That is, the lower the importance, the larger the pruning baseline coefficient, and the higher the pruning rate. Then, through... The function limits the calculated pruning rate to [ , Within the specified range, ensure that the pruning operation can effectively reduce the model complexity without significantly affecting the detection accuracy.

[0102] In summary, this layer-by-layer channel optimization strategy reduces the number of model parameters and computational load without significantly decreasing detection accuracy through a coherent process of "importance calculation - pruning rate allocation". First, a three-dimensional objective of "accuracy-parameter count-latency" is defined: using the validation set mAP@0.5 as a surrogate metric for accuracy, the number of parameters in each layer is counted, and the inference time of each layer is measured on the target deployment device (NVIDIA Jetson AGX Xavier). The three-dimensional objective data is normalized and linearly weighted to obtain the layer importance. Finally, the network layers are globally ranked according to importance, with high pruning rates assigned to shallow layers (low importance) and low pruning rates assigned to deep layers (high importance), and the pruning rate range is constrained by a truncation function.

[0103] The above-mentioned importance of the entire hierarchy The mathematical formula for quantitative calculation is as follows:

[0104] ;

[0105] ;

[0106] In the formula, , , These are the weighting coefficients. , , These represent the maximum values ​​of each indicator across all layers. Contribution to accuracy For parameters, This is for inference delay. To minimize pruning rate, To achieve the maximum pruning rate, This is a truncation function. This represents the final pruning rate.

[0107] By employing this layer-by-layer channel optimization strategy driven by layer importance, precise reduction of each layer of the YOLO11 model can be achieved. While ensuring that the detection accuracy mAP@0.5 of the coaxial cable connector decreases by no more than 2%, the number of model parameters is reduced by more than 30%, and the inference latency is reduced by more than 25%, significantly improving the feasibility of model deployment on industrial edge devices and real-time detection performance.

[0108] 3) Input the data to be detected into the coaxial cable connector detection model obtained in step 2) to accurately detect and locate the coaxial cable connector. Valid detection boxes are filtered using non-maximum suppression, and low-confidence prediction results are eliminated using a confidence threshold. Finally, a detection image labeled with the connector location and confidence level is output, completing the localization and identification of the coaxial cable connector. See the results below. Figure 4 As shown.

[0109] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A method for detecting coaxial cable joints based on an improved YOLO11 model, characterized in that, The improved YOLO11 model modifies the backbone network, feature fusion network, and model structure of the original YOLO11 model to achieve accurate detection and localization of coaxial cable connectors. Specifically, the backbone network is improved by replacing the downsampling module C2PSA with a high-efficiency channel attention module (EA) that features dynamic channel adaptation and small target feature enhancement, thereby improving the ability to focus on small target features. The feature fusion network is improved by embedding a lightweight edge enhancement module (LEE) based on a depthwise separable convolution module in the upsampling stage, thereby enhancing edge feature response. The model structure is improved by employing a layer-by-layer channel optimization strategy that differentiates by the importance of each layer. The importance value of each layer is calculated separately, and the number of channels in each layer is adjusted according to this importance value, achieving model lightweighting while maintaining accuracy. The specific implementation of the coaxial cable joint detection method includes the following steps: 1) Collect images of coaxial cable connectors from multiple devices and in multiple scenarios and perform preprocessing to obtain a coaxial cable connector target detection dataset with sufficient sample diversity; 2) The coaxial cable joint target detection dataset was input into the improved YOLO11 model for training. The optimal improved YOLO11 model with stable prediction accuracy was obtained as the coaxial cable joint detection model. During training, multi-scale features of coaxial cable joints were extracted by embedding an EA backbone network. The input feature map was first compressed and aggregated with channel information, then the dynamic receptive field module was fused to adjust the feature capture range, and finally attention weights were generated to strengthen key features. The feature fusion network with embedded LEE was used to optimize feature representation to extract edge information. The feature integrity was preserved by combining residual connections and activation functions, and then the channel dimension was matched by point convolution. At the same time, the model was lightweighted based on layer importance. The values ​​of three factors, including accuracy, number of parameters and latency, were quantitatively evaluated for each layer, i.e., the "accuracy-number of parameters-latency" ternary objective. The pruning rate was assigned according to importance to achieve channel pruning. 3) Input the data to be detected into the coaxial cable joint detection model obtained in step 2) to accurately detect and locate the coaxial cable joint. Then, use non-maximum suppression to filter valid detection boxes and combine confidence threshold to remove low confidence prediction results. Finally, output the detection image with the joint position and confidence level marked to complete the location and identification of the coaxial cable joint.

2. The coaxial cable joint detection method based on the improved YOLO11 model according to claim 1, characterized in that, Step 1) includes the following steps: Multi-device data acquisition: The core device is a D435i depth camera. During acquisition, the camera height is fixed, and the shooting angle covers the range of 0° to 45°, simultaneously acquiring RGB images and depth information; auxiliary devices include smartphones and web crawling; the acquisition scenarios cover 3 typical backgrounds and 5 light intensities of industrial assembly lines. Image preprocessing: Remove severely blurred, overexposed or underexposed images, and images without a target; then remove image noise by using a Gaussian filter with a convolution kernel size of 3×3, adjust image contrast by adaptive histogram equalization, and eliminate the influence of lighting differences by brightness normalization to ensure that the preprocessed image is clear and has a uniform grayscale distribution. High-precision image annotation: Use the LabelImg annotation tool to annotate the bounding box of the coaxial cable connector in YOLO format. The annotation rule is: the bounding box must completely enclose the connector area. Data augmentation: Various augmentation strategies are implemented on the labeled images to improve sample diversity and model generalization ability, ultimately resulting in a coaxial connector target detection dataset with sufficient sample diversity.

3. The coaxial cable joint detection method based on the improved YOLO11 model according to claim 1, characterized in that, The backbone network embedded with the EA performs the following operations: First, regarding the input dimension as The feature maps are subjected to global average pooling, where , For the feature map height and width, The total number of channels is determined by analyzing all spatial positions within each channel. The pixel values ​​are averaged. These represent the pixel values ​​on the x and y axes of the image, respectively, which were originally distributed across... Spatial dimension feature information is compressed and aggregated into A one-dimensional channel descriptor is used to effectively extract the global feature information of each channel and discard redundant details in the spatial dimension. Then, to avoid the surge in computational complexity and feature overfitting caused by an excessive number of channels, a 1×1 convolution operation is used to compress the number of channels in the one-dimensional channel descriptor to [a smaller number]. , / / represents integer division, The channel number compression design has two technical advantages: on the one hand, it achieves reasonable simplification of the number of channels through integer division, reducing the subsequent computational pressure; on the other hand, it forces the number of intermediate layer channels to be no less than 8 through the max function, ensuring that key feature information is not lost due to excessive dimensionality reduction during channel compression, thus ensuring the integrity of feature expression. Subsequently, a multi-level progressive activation mechanism is constructed, consisting of an unbiased linear layer, a memory-optimized ReLU activation function, and a sigmoid activation function. The unbiased linear layer design avoids interference from bias parameters on channel dependency learning. The memory-optimized ReLU activation function maintains non-linear feature extraction capabilities while achieving in-situ feature map computation through the `inplace=True` parameter setting, significantly reducing memory usage. Finally, the sigmoid activation function maps the output of the multi-level progressive activation mechanism to the [0,1] interval, generating channel attention weights that accurately reflect the importance of each channel. Finally, the generated channel attention weights are compared with the original input... The feature map is subjected to pixel-level multiplication operations on a channel-by-channel basis to weight and enhance the channels containing key features of the coaxial cable connector, while suppressing redundant information such as background noise and irrelevant textures, thereby improving the distinguishability between the target and the background in the feature map. The EA introduces a delayed fully connected layer initialization mechanism, which sets the fully connected layer parameter self.fc to None during the module initialization phase. Only when a specific input feature map is received during the first forward propagation does the adapted fully connected layer structure dynamically create based on the actual number of channels in the input feature map. This allows it to be integrated into variant networks of the YOLO11 model with different parameter scales such as n, s, m, l, and x, without needing to adjust the module structure for different network variants. The mathematical formulas corresponding to the entire operation process described above are as follows: ； ； ； ； ； In the formula, For the input feature map at location Number of channels pixel values, Number of channels The global average pooling result, This represents the feature vector after dimensionality reduction by an unbiased linear layer following the global pooling result. This indicates a linear layer without bias. Represents bias. Represents non-biased. For bias, This represents the result of global average pooling. This indicates that the number of output channels of the linear layer after dimensionality reduction is a fixed value. , The number of output channels of a linear layer is equal to the total number of channels of the input feature map. , This represents the memory-optimized ReLU activation function. The attention weights of the channels, For the first Attention weights for each channel, For in position Number of channels The output feature map after weight enhancement; Through the aforementioned backbone network, four downsampling operations are performed sequentially during feature extraction, with downsampling rates of 2, 4, 8, and 16, respectively. By gradually increasing the channel dimension of the feature map and reducing the spatial resolution, a progressive extraction from shallow texture features to deep semantic features is achieved, ultimately resulting in four levels of output feature maps from shallow to deep, denoted as P3, P4, P5, and P6, respectively. Among them, P3 has the highest spatial resolution, which can accurately capture the subtle textures and edge details of the coaxial cable connector, while P4 enhances the semantic feature expression while preserving the spatial resolution. The two work together to optimize the feature extraction needs of small targets such as coaxial cable connectors.

4. The coaxial cable joint detection method based on the improved YOLO11 model according to claim 1, characterized in that, The feature fusion network embedded with LEE adopts the original structural framework of the original YOLO11 model to ensure the compatibility of the network structure and basic detection performance. The core workflow of LEE consists of three closely linked technical steps, which work together to achieve efficient extraction and optimization of edge features, as detailed below: The first step is efficient edge feature extraction, which adopts an innovative design using a depthwise separable convolutional architecture. This completely eliminates the redundant computations in the feature extraction process of traditional standard convolutions. First, the input feature map is processed through a 3×3 depthwise convolutional layer. The convolution kernel of this 3×3 depthwise convolutional layer corresponds one-to-one with the number of channels in the feature map. That is, each convolution kernel performs spatial dimension convolution operations only within a single channel, which can accurately focus on the spatial position information and texture features of the coaxial cable joint edge. After completing the single-channel spatial edge feature extraction, cross-channel feature fusion is performed through a 1×1 point convolutional layer. This 1×1 point convolutional layer can integrate the edge features extracted from each channel individually, and flexibly adjust the channel dimension of the feature map according to the needs of subsequent feature processing, so as to achieve cross-channel complementarity and dimension adaptation of edge features. The second stage is dynamic adaptive channel adjustment, which innovatively introduces a dual-delay initialization mechanism. That is, the 1×1 point convolutional layer and the 3×3 depth convolutional layer in the module do not predefine the specific number of input and output channels and convolutional kernel parameters during the module initialization stage. Instead, when the module performs its first forward propagation, it dynamically creates an adapted convolutional layer structure and convolutional kernel parameters based on the channel dimension and spatial size of the actual received input feature map. This completely breaks the fixed limitation of the number of input feature map channels in the traditional edge enhancement module, enabling it to flexibly adapt to the multi-scale model of YOLO11 model with different parameter scales such as n, s, m, l, and x without any structural modification. The third step is feature integration and performance optimization. It uses dual technical means to ensure the integrity of features and the efficiency of model operation during the edge enhancement process. On the one hand, a residual connection mechanism is introduced to directly add and fuse the edge features extracted by depthwise separable convolution with the original input features of the module. This can effectively avoid the loss of basic feature information during the edge feature enhancement process and ensure that the feature map contains both accurate edge details and retains the complete overall features of the target. On the other hand, the SiLU activation function is used to replace the traditional ReLU activation function. The SiLU activation function has smooth nonlinear characteristics, which can enhance the nonlinear expression of features while optimizing the gradient flow of the model.

5. The coaxial cable joint detection method based on the improved YOLO11 model according to claim 4, characterized in that, The LEE performs the following operations: For feature maps of arbitrary dimensions input to the feature fusion network First, spatial edge features are extracted using a dynamically initialized 3×3 depth convolutional layer. This 3×3 depth convolutional layer is based on... The actual number of channels dynamically generates a corresponding number of 3×3 convolutional kernels, performing convolution operations only within a single channel to accurately capture key edge features such as the port contour and metal seam of the coaxial cable connector, while avoiding redundant cross-channel computations in traditional convolution. Next, a dynamically initialized 1×1 point convolutional layer is used to perform cross-channel fusion of the edge feature maps extracted by depth convolution, integrating edge information from different channels and initially adjusting the channel dimensions of the feature maps. Finally, a residual connection mechanism is activated to combine the edge features processed by point convolution with the original input features of the module. Pixel-by-pixel addition is performed to obtain the feature map after residual fusion. Ensure that essential features are not lost; Then The input is fed into the SiLU activation function, whose non-linear transformation enhances the discriminative power of the features, further strengthening the difference between the coaxial cable connector edge features and the background. Finally, another dynamically initialized 1×1 point convolutional layer is used to fine-tune the number of channels in the activated feature map, ensuring the final output feature map... The input dimensions of the feature fusion network are fully matched with those of the subsequent feature fusion layers, ensuring the smoothness of the entire feature fusion network in the feature processing flow. The mathematical formulas corresponding to the entire operation process described above are as follows: ； ； ； ； ； In the formula, A 3×3 depthwise convolution kernel dynamically initialized for each channel. Where is the row index of the convolution kernel. The column index of the convolution kernel. The input channel index bound to the convolution kernel; To extract the spatial location x-coordinate of depthwise convolution. The vertical axis is The number of channels is Edge feature map; The spatial location of the input feature map; This is the output feature map of the first 1×1 pointwise convolution; , These are dynamically initialized 1×1 point convolution weight matrices, the former used to fuse channel information and the latter used to match the channel dimensions of subsequent layers. This is the edge enhancement feature map after SiLU activation; Use the Sigmoid activation function; This is the feature map after residual fusion. This is the final output edge enhancement feature map.

6. The coaxial cable joint detection method based on the improved YOLO11 model according to claim 1, characterized in that, The layer-by-layer channel optimization strategy does not employ traditional uniform or random pruning methods. Instead, it intelligently allocates the pruning rate for each layer through hierarchical importance quantification calculation, and constrains the pruning range. This process achieves precise reduction of model parameters and computational load while maximizing the preservation of model detection accuracy. The implementation steps of this layer-by-layer channel optimization strategy are as follows: The first step is to perform hierarchical metric calculations. First, we clarify three optimization objectives: accuracy, number of parameters, and latency. This covers the core evaluation dimensions of model performance, avoiding performance imbalances caused by optimizing a single dimension. Specifically, we use the value mAP@0.5 from the dataset. This value represents the average accuracy when the intersection-over-union (IoU) threshold is 0.5, serving as a core proxy metric for model accuracy. mAP@0.5 directly reflects the model's precision and recall capabilities for detecting coaxial cable joints and is a recognized key accuracy evaluation standard in object detection tasks. By traversing the parameter matrices of each layer of the model, we count the number of parameters in each layer, including learnable parameters such as convolution kernel parameters and bias parameters, as a core metric for measuring the computational complexity of that layer. Finally, we conduct actual inference tests on the NVIDIA Jetson AGX Xavier edge computing platform, recording the individual inference time of each network layer using a high-precision timer as a core metric for measuring the real-time performance of that layer. After acquiring the ternary target data for each layer, normalization is performed on each type of indicator to eliminate the influence of differences in dimensions between different indicators. The specific normalization method is as follows: for precision indicators... , representing the first The accuracy contribution value corresponding to the layer was obtained by testing using the layer-by-layer masking method, and then mapped to the [0,1] interval using min-max normalization; for parameter indicators, the first... Layer parameter quantity Similarly, min-max normalization is used to map to the [0,1] interval; for the latency metric, the first... Layer inference time The reverse min-max normalization is used to map to the [0,1] interval. The purpose of reverse normalization is to make the layer with shorter latency get a higher normalization score, which is consistent with the optimization direction of accuracy and parameter quantity indicators. Subsequently, to balance the importance of the ternary objective data, weighting coefficients α, β, and γ are introduced, where α, β, and γ are preset non-negative weight values ​​based on actual needs, satisfying α + β + γ = 1. The normalized ternary objective data are then linearly weighted and summed to obtain the hierarchical importance of each level. Its calculation formula is The design logic of this formula is: the layer that contributes more to accuracy, The larger the value, the fewer the number of parameters in the layer. The larger the value, The larger the value, the shorter the inference time of the network layer. The larger the value, The larger the value, the more accurately the overall importance of each layer can be quantified through weighted summation. The above-mentioned importance of the entire hierarchy The mathematical formula for quantitative calculation is as follows: ； In the formula, , , These represent the maximum values ​​of each indicator across all layers. Contribution to accuracy For parameters, For inference delay; The second step is intelligent allocation of differentiated pruning rates, after obtaining the hierarchical importance of all layers. Then, apply the following to all layers: The values ​​are globally sorted from largest to smallest to clarify the importance level of each layer in the entire model. Based on the sorting results, a differentiated pruning strategy is implemented: for shallow networks with low layer importance, which are responsible for extracting basic texture features and have many redundant parameters, but have little impact on the final detection accuracy, a high pruning rate is assigned, that is, their number of channels is significantly reduced to minimize the computational cost of the model; for deep networks with high layer importance, which are responsible for extracting high-level semantic features and play a decisive role in the target detection accuracy, they have few redundant parameters, a low pruning rate is assigned, and only a small number of redundant channels are reduced to retain their feature extraction capabilities to the maximum extent. The third step is to constrain the pruning range. To avoid a severe decrease in feature extraction capability due to an excessively high pruning rate, or an insignificant lightweighting effect due to an excessively low pruning rate, a cutoff function is introduced. Limit the pruning rate of each layer to a range and set a minimum pruning rate. With the maximum value The final calculation of the pruning rate is done using the formula. Implementation, in which For the final pruning rate, This is the pruning intensity adjustment coefficient, used to control the overall aggressiveness of pruning. This represents the pruning baseline coefficient, which is inversely proportional to the importance of the hierarchy. That is, the lower the importance, the larger the pruning baseline coefficient, and the higher the pruning rate. Then, through... The function limits the calculated pruning rate to [ , Within the specified range, ensure that the pruning operation can effectively reduce the model complexity without significantly affecting the detection accuracy.

Images