Research on the Image Recognition Framework for Drone-Based Inspection of Photovoltaic Plants Powered by Deep Learning

Abstract

This paper proposes a deep learning-based image recognition framework for UAV-based photovoltaic power plant inspections to address the low efficiency of traditional manual inspections and the inaccuracy of existing recognition models. This framework utilizes a three-stage architecture: preprocessing, feature extraction, and recognition and classification. It optimizes image quality through adaptive median filtering and a multi-scale Retinex algorithm. It enhances feature extraction capabilities using a modified YOLOv8 network and integrates multi-task learning to achieve accurate defect recognition. Experiments using a dataset of 10,000 images (covering eight defect categories) show that the framework achieves an average accuracy of 96.8%, a precision of 95.2%, a recall of 94.5%, and an F1 score of 94.8%. This represents an 8.3% improvement in accuracy over VGG16 and a 5.2%-6.7% improvement in F1 score for small object defects over YOLOv5. The processing speed reaches 60 FPS. The ROC curve achieved an AUC value of 0.986, and the confusion matrix diagonal elements accounted for over 94%, demonstrating the framework's high accuracy and stability, providing strong support for intelligent operation and maintenance of photovoltaic power plants.