CCA-YOLO: Channel and Coordinate Aware-Based YOLO for Photovoltaic Cell Defect Detection in Electroluminescence Images

Solar energy is a renewable energy used for urban power generation, contributing to sustainable cities. In solar energy generation, it is important to inspect the health of photovoltaic (PV) cells for safety and power transformation efficiency. Defects in PV cells are usually irregular with different scales, challenging automated defect detection for PV cells. Therefore, this article presents a channel and coordinate aware-based YOLO (CCA-YOLO) for efficient PV cell defect detection. Specifically, to provide accurate backbone features from the complex background defect images, the residual coordinate convolution-based ECA (RCC-ECA) enhances the backbone feature representation by learning from channel and coordinate information. To learn the intraclass/interclass variations and interclass similarity and convey coordinate information among different scales, the multiscale defect feature localization module (MDFLM) incorporates a larger backbone feature to improve the robustness of multiscale defects. The RCC-Up/Down optimizes the sampled features to minimize the inaccurate representation of the features caused by the sampling process. In addition, RCC-Up/Down conveys the coordinate information during the up/down sampling process to maintain coordinate awareness, which allows the network to learn from the coordinate information efficiently. Furthermore, the residual feature fusion with coordinate convolution-based CBAM (RFC-CBAM) is introduced to maintain the channel and coordinate awareness for efficient learning from fused features. The proposed CCA-YOLO outperforms state-of-the-art (SOTA) methods in PVEL-AD on precision (71.71%), recall (76.91%), F1-Scores (74.19%), mAP50 (98.57%), $\text {AP}_{S}$ (26.80%), $\text {AP}_{M}$ (64.78%), and $\text {AP}_{L}$ (74.93%).