To address issues of low detection accuracy and high false-positive and false-negative rates in solar cell defect detection, this paper proposes an optimized solar cell electroluminescent (EL) defect detection model based on the YOLOv8 deep learning framework. First, a self-calibrated illumination (SCI) method is applied to preprocess low-light images, enhancing effective feature information for solar cell defects. Then, a space-to-depth (SPD) attention module is introduced, replacing the second stride convolution layer in the backbone network. This substitution avoids information loss caused by stride convolution, expands the receptive field, and reduces computational load, preserving more feature information during extraction. Next, a spatial-BiFPN (S-BFPN) network is constructed to perform multi-scale feature fusion, stabilizing defect recognition rates by addressing the shape variability of solar cell defects. Lastly, the loss function is improved by adopting MPDIoU, which resolves the issue of ineffective penalties in the original CIoU loss function. The experimental results show that the improved YOLOv8 model achieved an mAP of 96.9%, a 2.2% increase compared to the original YOLOv8. The computational load was reduced by 0.2 GFlops, and the detection speed reached a maximum of 155 f/s, demonstrating high accuracy and real-time performance, making it more suitable for industrial deployment.