A lightweight semantic segmentation model Faster DeeplabV3 + based on DeeplabV3 + network was improved to address the issues of large parameter count, slow computation speed, and low efficiency in traditional semantic segmentation models. The Faster DeeplabV3 + model used lightweight MobilenetV2 instead of Xception as the backbone feature extraction network, significantly reducing the number of parameters and improving computational speed. Using the combination of deep separable convolution (DSC) and air space pyramid expansion convolution (ASPP) , a new depth separable dilated convolution(DSD-Conv) was developed, namely, the depth separable empty space pyramid module (DP-ASPP) , which expanded the receptive field while reducing the original convolution parameters and facilitating faster operation. In order to enhance the sensitivity and accuracy of the network in extracting feature information from different dimensions, a dual attention mechanism module was added for processing low-level and high-level feature maps generated in the coding region. Two loss functions, cross entropy and Dice Loss, were integrated to provide more comprehensive and diverse optimization for the model. The improved model was tested on the PASCAL VOC 2012 dataset. The experimental results show that the average intersection to union ratio increases from 76.57% to 79.07%, and the segmentation accuracy increases from 91.2% to 94.3%. The network parameter count (parameters) of the improved model is reduced by 3.86 ×10⁶, and the floating-point computational load (GFLOPs) is decreased by 117.98 G. The proposed algorithm significantly reduces parameter count and improves computational speed, while also optimizing the segmentation effect.