Research On Insulator Defect Detection Algorithm Based On Multi-feature Enhancement And Improved YOLOv5

With the development of smart grid,the detection of component faults increasingly relies on computer,network,and artificial intelligence technologies.In power transmission systems,faults in the components that connect the lines can cause extensive power grid failures.Therefore,it is essential to improve the accuracy of fault diagnosis and the timely maintenance and replacement of these components.However,due to the complex environmental backgrounds and various weather conditions,the error rate of fault detection based on image collection is relatively high,especially for components such as insulators in transmission lines.To address this issue,this thesis proposes to optimize image data and further consider the actual environment during collection,and to use deep neural networks for defect detection.The main contributions of this work are as follows:(1)An improved automatic color balance algorithm is proposed.To address the challenges of complex image backgrounds,noise,and low resolution in the original dataset,the paper focuses on image denoising and enhancement algorithms.After comparing multiple denoising algorithms,it is found that using the fast matching 3D filtering algorithm provides the best denoising results,but it is time-consuming and computationally complex.Finally,the paper uses the pre-trained deep learning weight Noise2 Noise to perform batch denoising on the images.In addition,an optimized automatic color balance algorithm is proposed,which introduces logarithmic transformation,white balance,and feature pyramid algorithms to improve the algorithm’s operating speed while achieving good experimental results.(2)A mosaic data augmentation algorithm and an image fogging algorithm are constructed.To address the problem of insufficient training data,the thesis uses an atmospheric light model to fog the images,thereby expanding the dataset while making it more consistent with the environment in which insulator images are actually collected.Finally,the paper constructs feature point detection,image segmentation algorithms,and morphological processing to analyze and locate the image features of insulator self-burst defects.(3)A lightweight improved YOLOv5 algorithm is proposed.To address the problem of imbalanced positive and negative samples in the dataset,the paper first introduces the EIOU strategy to alleviate the problem of sample imbalance,using electric visual technology.Secondly,the paper proposes an improved network structure,which introduces a lightweight Ghost Net module and NAM(Normalization-based Attention Module)attention mechanism,reducing the model parameter size while maintaining similar accuracy.Finally,the proposed algorithm achieves final m AP results of 97.9% and 96.5% for insulator detection and self-burst defect detection,respectively,which are better than those of other models.