Abstract: Aimed at the commonly used computer vision-based photovoltaic (PV) panel defect recognition model to extract difficult features, slow inference speed, and other problems, this paper proposes a YOLOv8 improvement algorithm that incorporates a two-layer routing attention mechanism visual converter biformer as the backbone network, which improves the model adaptability by dynamically adjusting the attention span. A feature extraction network is designed incorporating the squeeze-and-excitation attention mechanism, which improves the sensitivity of the network to key features, thus improving the feature characterization. The better student semi-supervised learning method (BSLM) is proposed, which improves the model accuracy while reducing the amount of data labeling. Experimentally, it was proven that the mean average precision at the intersection over union threshold of 0.5 (mAP@0.5) of the proposed algorithm reached 83.9%, which is 3.1% higher than that of the YOLOv8n algorithm, far exceeding other one-stage algorithms. It can also better meet the accuracy requirements of defect detection in PV panels. The frames per second reached 101.01, which satisfies the real-time requirements of defect detection in PV panels. After using the BSLM proposed in this paper, the mAP@0.5 reached 90.7%, which further improved the model performance.