An Enhanced Solar Panel Crack Detection System Using Resource Optimized Depthwise Convolutional Feature Incorporated Vision Transformer Model

Abstract

The defects in solar panels reduce the solar energy production and hence the detection of these defects in solar panels is important. The timely intervention and inspection of the cells in the solar panels reduce the losses in the solar panels energy production process. This can be done by the automated solar panel inspection method. This research work proposes a fully automated Artificial Intelligence (AI) based solar panel crack detection system using Convolutional Neural Network (CNN) feature incorporated with Vision Transformer Model (VTM). The main objective of this work is to classify the drone captured solar panel images into either cracked or non-cracked. This proposed work is designed with the Anisotropic Diffusion (AD) filter, Multi-Window Gabor Transform (MWGT), Resource Optimized Depthwise Convolutional Neural Networks (RODCNN) architecture and Vision Transformer Model (VTM). AD filter is used to smooth the regions while suppressing noises in images. The MWGT enhances the time-frequency resolution property of each pixel in the AD filtered image and then a novel RODCNN architecture is designed for computing the solar panel features. The features are further classified using the VTM classification approach, which classifies the solar panel images into either cracked or non-cracked. This proposed methodology is tested and validated on two solar panel imaging datasets RoboFlow and Kaggle.The average computational crack detection results have been attained as 98.84% SPSe, 99% SPSp, 98.94% SPA, 98.97% SPPR, 99.07% FSM, and 99.07% MCC in RoboFlow dataset. The average computational crack detection results have been attained as 98.93% SPSe, 99.01% SPSp, 98.96% SPA, 98.98% SPPR, 98.99% FSM, and 98.91% MCC in Kaggle dataset. The key values of the experimentations of the proposed methodology are compared with the recent state-of-the-art methods.