Deep learning-based method for defect detection in electroluminescent images of polycrystalline silicon solar cells

Abstract

To achieve defect detection in bare polycrystalline silicon solar cells under electroluminescence (EL) conditions, we have proposed ASDD-Net, a deep learning algorithm evaluated offline on EL images. The model integrates strategies such as downsampling adjustment, feature fusion optimization, and detection head improvement. The ASDD-Net utilizes the Space to Depth (SPD) module to effectively extract edge and fine-grained information. The proposed Enhanced Cross-Stage Partial Network Fusion (EC2f) and Hybrid Attention CSP Net (HAC3) modules are placed at different positions to enhance feature extraction capability and improve feature fusion effects, thereby enhancing the model's ability to perceive defects of different sizes and shapes. Furthermore, placing the MobileViT_CA module before the second detection head balances global and local information perception, further enhancing the performance of the detection heads. The experimental results show that the ASDD-Net model achieves a mAP value of 88.81% on the publicly available PVEL-AD dataset, and the detection performance is better than the current SOTA model. The experimental results on the ELPV and NEU-DET datasets verify that the model has some generalization ability. Moreover, the proposed model achieves a processing frame rate of 69 frames per second, meeting the real-time defect detection requirements for solar cell surface defects.