Defect detection of photovoltaic modules based on improved SSD algorithm

• Published in: 2023 International Conference on Image Processing, Computer Vision and Machine Learning (ICICML) pp. 1063 - 1066
• Main Authors: Kong, Xiangyu, Xu, Wanxuan, Xu, Bohao, Jin, Hui, Zhan, Peihong
• Format: Conference Proceeding
• Language: English
• Published: IEEE 03.11.2023
• Subjects: Attention mechanism; Classification algorithms; Deep learning; Image recognition; Inspection; Machine learning algorithms; Maintenance engineering; Photovoltaic detection; Photovoltaic systems; SSD algorithm; Transfer learning
• DOI: 10.1109/ICICML60161.2023.10424859

Automated inspection of unmanned aerial vehicles (UAVs) offers an effective solution to address the operational and maintenance requirements of large-scale distributed photovoltaic systems. Due to the abundance of image data generated by UAVs, a robust algorithm is needed to achieve higher recognition accuracy and faster processing speed. In this regard, we propose an enhanced Single Shot MultiBox Detector (SSD) algorithm for detecting defects in photovoltaic (PV) modules. The novel algorithm incorporates an attention mechanism into the original SSD algorithm and leverages transfer learning to enhance both detection speed and accuracy. It is capable of automatically identifying and classifying common PV module defects such as glass breakage, yellowing of the light receiving surface, and dust accumulation. Experimental results demonstrate that the improved SSD algorithm outperforms other state-of-the-art algorithms including Faster-RCNN, YOLO3, and VGG16-SSD in terms of recognition accuracy, recall rate, and detection speed. Consequently, the proposed algorithm significantly enhances the efficiency of defect recognition in photovoltaic modules.