Research On Small Target Detection And Rust Defect Recognition Method Based On Transmission Line Clamp

With the rapid pace of urbanization,there has been an increase in demand for electricity,and the inspection of transmission lines is crucial to ensuring stable power operation.With the advancement of UAV technology and deep learning,power inspection has benefitted from the introduction of UAV detection and computer vision algorithms.The wire clamp,a significant metal appliance on the transmission line,plays a vital role in maintaining the line’s proper functioning.In the event that the clamp rusts and falls off,it can severely impact the line’s operation and potentially lead to electric accidents.As a result,the inspection of wire clamps is of utmost importance and must be given top priority.Currently,there are two critical issues that hinder UAV inspections for detecting transmission line clamp defects.Firstly,the proportion of line clamps captured in aerial images is minimal,which leads to numerous missed detection problems with conventional detectors.Secondly,the rusted and unrusted datasets of wire clamps suffer from a severe long-tailed imbalance problem.As a result,the classifier’s prediction results tend to favor the category with a large number of samples,as accelerated by the cross-entropy loss.This paper presents a novel Top-Down detection and identification approach for UAV inspection of transmission lines.The proposed method utilizes an enhanced YOLOv5 framework for wire clamp detection and delves into the defect recognition task,proposing refined Res Net-50 and TIN algorithms.The research objectives of this study are outlined as follows:(1)The proposed Top-Down detection and recognition scheme involves three key steps.Firstly,a larger area fitting image is detected from the aerial image.Subsequently,the clamp image is detected from the fitting image to determine its precise location.Finally,the clamp image is forwarded to the classifier for defect detection and identification.(2)Designed an improved target detection model based on the YOLOv5 l framework.In response to the insufficient feature extraction and limited multi-scale feature learning of the YOLOv5 l Backbone,we proposed the use of a multi-layer cascade module and over-sampling module to replace the C3 module and convolutional down-sampling layer in the Backbone.Additionally,the Pyramid Compression Attention(PSA)method was introduced to enable the model to learn features at different scales,thus enhancing the detection of small objects.In the network Neck section,we extended the CSPNet and pyramid concepts,resulting in the creation of the CSPSPP module to replace the original SPPF.Moreover,the C3 module and convolutional downsampling layer in Neck were substituted with a multi-layer cascade module and a downsampling module.To improve the model’s ability to predict small targets,we introduced the SIo U Loss,which replaced the original CIo U positioning loss in the loss part.Finally,we conducted verification experiments to compare the model series scheme and the direct detection scheme,demonstrating the effectiveness of the improved model and detection scheme presented in this paper.(3)Proposed an improved Res Net-50 model and TIN algorithm for line clamp defect classification.The improved Res Net-50 model uses pixel attention to highlight the feature differences between rusted and unrusted images,and adopts a resampling scheme to undersample the unrusted category to reduce the sample quantity difference.We also use data augmentation and random erasing strategies to expand the sample size and help the model recognize occluded line clamps.Finally,we use Focal Loss to improve the accuracy of the rusted category.Although this approach has made significant progress,the accuracy of the rusted category is still not sufficient for engineering requirements.Therefore,we propose the TIN model,an end-to-end multi-module network that combines visual supervision and text supervision information.We compare and learn from a CNN visual module and a Transformer text module,then use the Large Kernel Convolution Attention(LKA)to further learn spatial and channel information,and adopt Seesaw Loss to suppress the negative sample gradients of the rusted category.Unlike mainstream long-tail models,TIN does not require additional finetuning or secondary training of the dataset to significantly improve the accuracy of the rusted category and meet engineering requirements.Finally,we demonstrate the advantages of TIN through extensive comparative experiments and connect the previously proposed detection model and classification model using a Top-Down approach to verify the effectiveness of the proposed model and recognition approach.(4)To address the memory and parameter-heavy nature of the detection model,this study proposes a lightweight structure design.Depthwise separable convolution replaces the 3× 3standard convolution in the detection model,enabling a more efficient network design.To optimize the inference speed,the combined network layer and half-precision FP16 in Tensor RT are utilized.Finally,the model is deployed on the server for practical use.Building on the aforementioned research,this paper successfully achieves the accurate detection of transmission line clamps and identification of rust defects.This solution for transmission line inspection and intelligent operation and maintenance offers significant engineering value and research significance.