HaarSSFF-YOLOv8: a deep convolutional neural network for photovoltaic defect

This paper presents HaarSSFF-YOLOv8, a photovoltaic defect detection algorithm based on YOLOv8.The proposed algorithm tackles key challenges in photovoltaic electroluminescence defect detection, including diverse defect sizes, similar defect morphologies, and the detection of small or sparse defects. The algorithm introduces several novel enhancements to improve detection performance. First, the neck module of YOLOv8 is replaced with Haar wavelet convolution, dynamic max pooling, and dynamic average pooling, enabling more effective feature extraction. This modification improves global feature information fusion by enhancing long-range feature concatenation, thereby preserving rich edge and texture features. Second, a channel-spatial synergistic attention mechanism is incorporated into the bottleneck structure to reduce redundant features and optimize feature representation. Finally, the integration of the SPPFCSPC module facilitates the extraction of high-level semantic information. Compared to the baseline YOLOv8 model, the improved algorithm achieves significant performance gains, with [email protected] and [email protected] increasing by 2.2% and 3.8%, respectively, on the PASCAL VOC2007 dataset. On a photovoltaic defect dataset, the proposed model achieves impressive [email protected] and [email protected] scores of 94.7% and 71.3%, respectively. The experimental results demonstrate that replacing the bottleneck structure's downsampling modules with Haar wavelet convolution, dynamic max pooling, and dynamic average pooling enriches the network's low-level semantic information. This innovative approach significantly enhances the model's ability to capture intricate edge and texture features, addressing key limitations in photovoltaic defect detection and improving overall detection accuracy.