Deeplab-YOLO: a method for detecting hot-spot defects in infrared image PV panels by combining segmentation and detection

• Published in: Journal of real-time image processing Vol. 21; no. 2; p. 52
• Main Authors: Lei, Ye, Wang, Xiaoye, An, Aimin, Guan, Haijiao
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01-04-2024; Springer Nature B.V
• Subjects: Accuracy; Algorithms; Classification; Computer Graphics; Computer networks; Computer Science; Datasets; Defects; Energy resources; Image processing; Image Processing and Computer Vision; Image segmentation; Infrared imagery; Model accuracy; Multimedia Information Systems; Panels; Pattern Recognition; Photovoltaic cells; Power plants; Semantics; Signal,Image and Speech Processing; Deep learning; Deeplabv3; YOLO v5; Infrared image of PV panel; Defect detection
• ISSN: 1861-8200; 1861-8219
• DOI: 10.1007/s11554-024-01415-x

Aiming at the problem of difficult operation and maintenance of PV power plants in complex backgrounds and combined with image processing technology, a method for detecting hot spot defects in infrared image PV panels that combines segmentation and detection, Deeplab-YOLO, is proposed. In the PV panel segmentation stage, MobileNetV2 was introduced into the Deeplabv3+ model. Empty convolution in the atrous spatial pyramid pooling (ASPP) structure was improved, and established a relationship between layer-level features, the CBAM attention mechanism was combined, which achieved fast and accurate segment of PV panels and avoided false detection of hot spots. In the hot-spot recognition stage, a lightweight MobileNetV3 network was designed to replace the YOLO v5 backbone network, a small defect prediction head was added, and EIOU was used as a loss function, which improved the speed and accuracy of hot-spot detection and enhanced the performance of the YOLO v5 model. The experimental results show that the optimized Deeplabv3+ model and YOLO v5 model improve the accuracy of segmenting PV panels in images and identifying hot-spot defects by 2.61% and 0.7%, respectively, compared with the original model. This proposed method can accurately segment the PV panels and then identify different sizes of hot-spot defects on the PV panels.