Abstract: To solve the segmentation problem of infrared images of photovoltaic modules,MobileNetv2 is used as the backbone feature extraction network of DeepLabv3+ and the location channel attention module is used to reduce background interference. Mixed strip pooling is introduced to optimize the ASPP module,which helps the model to further capture global and contextual information. The DeepLabv3-T network is designed for rooftop PV modules with difficult detection. Based on the above improvements,texture information is incorporated into the selective background suppression to achieve accurate segmentation of PV modules. Experimental results on the PV＿large and PV＿roof datasets demonstrate that the text-based approach is superior to the prior art,and the mIoU value of deeplabv3-t is 2.74% and 7.93% higher than that of DeepLabv3+,respectively. In addition,ablation experiments are designed to demonstrate the effectiveness of each improved module.