Abstract: To address the issue of low detection efficiency of surface defects for reconstituted tobacco, a detection method based on a lightweight semantic segmentation model was proposed and evaluated. Firstly, reconstituted tobacco samples were batch processed to construct a sample dataset. Secondly, the Labelme software was used to label and classify the defects, and sufficient model training samples were obtained through the sliding window sampling method. Finally, the lightweight semantic segmentation model Deeplab-T was constructed using MobileNetV2 and the void convolution as the backbone network, and the convolutional block attention module was introduced in the feature fusion process to highlight the segmentation target. The results showed that: 1) The mean intersection over union and mean pixel accuracy of Deeplab-T model were 76.35% and 83.71%, respectively, which were 13.98 and 15.32 percentage points higher than those of the DeeplabV3+ model. Additionally, the detection frames per second was increased by 234.10%, meeting the industrial requirements for both accuracy and speed. 2) The Deeplab-T model achieved the contour segmentation of reconstituted tobacco defects in different directions, proportions and of different types with good robustness. This study provides a theoretical reference for the online monitoring of typical defects in reconstituted base sheet making process.