Abstract: Aiming at the problems of high computational complexity, large number of parameters, slow segmentation speed and low segmentation accuracy existing in the photovoltaic cell defect segmentation model, a photovoltaic cell defect segmentation model based on lightweight improved U-Net is proposed. First of all, the MobitNetv3_Large network is used to replace the backbone network of the original U-Net, which reduces the computational amount and the number of parameters of the model while retaining the feature extraction ability of the original network. Secondly, the G-DConv module is designed by integrating the DynamicConv module into the GhostConv module, replacing the ordinary convolutional module used in the upsampling part of the original U-Net, which maximally reduces the network parameters and computational amount while improving the inference speed of the model. Finally, by introducing the ECA attention mechanism after network upsampling, the interference of complex background on the detection effect is reduced. The experimental results show that the number of parameters of this model is only 2.43×10⁶, the computational amount is only 3.03×10⁹, and the inference speed reaches 61 frame/s. Compared with the baseline model, the improved model increases MIoU and MPA by 0.12 and 2.17 percentage points respectively, meeting the requirements for industrial equipment deployment.

Key words: photovoltaic cell, U-Net, light weight, semantic segmentation, ECA