To address the challenges of mutual occlusion and accurate differentiation between rice and weeds in real-world environments, an improved method for rice-weed recognition was proposed based on DeepLabv3+. The research focused on images of rice field weeds captured by UAV in complex backgrounds, the MobileNetv2 was used as the backbone feature extraction network to reduce the number of parameters and computational complexity of the model; channel and spatial dual-domain attention modules were integrated to strengthen the model's attention to important features. A multi-branch receptive field cascade fusion structure was proposed based on dense sampling to improve the ASPP module to expand the sampling range. In addition, improvements to the decoder were made. Experimental results demonstrated that the improved model achieved the best performance in rice-weed recognition, with a mean intersection over union (MIoU) of 90.72%, mean pixel accuracy (mPA) of 95.67%, and F1_score of 94.29%, which were 3.22, 1.25, and 2.65 percentage points higher than that of the basic model. The improved model had a size of 11.15MB, which was 1/19 of the original model's size, and achieved an average network inference speed of 103.91 frames per second per image. The results demonstrated that the improved model can accurately segment rice and weeds in complex backgrounds, supporting precise pesticide application using UAV.