Aiming at the problems of weak feature extraction ability,low segmentation accuracy and missing segmentation in solar cell defect segmentation,an improved U²-net solar cell defect segmentation method is proposed. To improve the extraction ability of effective features in RSU and reduce the number of parameters,the residual structure is used to combine the effective channel attention module and the depth separable convolution to form a new feature extraction layer.In order to prevent the loss of spatial information,a semantic embedded branch structure is added to the outer codec hop connection,and CARAFE operator is used for upsampling to introduce more semantic information intolow-level features to strengthen the fusion of features between levels,and reduce the spatial information lost due to jump connection.Finally,the proposed method is compared with the commonly used segmentation network. The experimental results show that the classification pixel accuracy,IOU and MIOU of this method are 74.69%, 60.68%and 80.30%respectively.Compared with U-Net,PSPNet and Deeplab v3+,this method not only effectively improves the accuracy of defect segmentation,but also realizes the accurate segmentation of small target defects and effectively reduces missing segmentation