YOLO-iCBAM: an improved YOLOv4 based on CBAM for defect detection

Defect detection in Photovoltaic (PV) cell Electroluminescence (EL) images is a challenge in industry. In this paper, a novel defect detection method YOLOv4 with an improved Convolutional Block Attention Module (YOLO-iCBAM) is proposed for PV cell EL images. We first propose an improved CBAM to enhance the network's ability to capture multi-scale defects in complex image backgrounds. Then, we modify the conventional YOLOv4 architecture for defect detection. Specifically, we adjust the backbone network to make a fast convergence. Then, we adopt the iCBAM to YOLOv4 to refine the feature map before YOLO Head. Then, we train a K-Means++ model based on PV cell EL images to generate anchors for bounding box regression. Moreover, we conduct experiments in the PVEL-AD dataset to evaluate the proposed YOLO-iCBAM. The experimental results indicated that the proposed YOLO-iCBAM achieves a better F1-Score of 0.716 and mAP of 0.748.