Lightweight Semantic Segmentation Network based on Feature Fusion and Attention Mechanism

To address the issue of excessive parameter count, as well as the poor correlation among multi-scale feature maps and feature maps from different stages in the DeepLabv3+ network, a lightweight semantic segmentation network is proposed by introducing feature fusion and attention mechanisms on the basis of the DeepLabv3+ model. Firstly, the backbone network of the original DeepLabv3+ model is replaced with the optimized MobileNetv2 network to reduce model complexity. Secondly, an unconventional receptive field fusion strategy is employed to enhance the ASPP structure, thereby improving the information utilization and training efficiency of ASPP. Different dilation rates in convolutions are followed by the introduction of the CAM attention mechanism. Subsequently, the SAM attention is introduced after feature fusion, and the low-dimensional features are incorporated with the CBAM attention mechanism to focus on image edge features, thereby enhancing feature extraction accuracy. Finally, the model is trained using the focal loss function to improve segmentation accuracy. Experimental validation is conducted on the PASCAL VOC 2012 dataset, and the results indicate that the proposed approach further enhances segmentation performance while significantly reducing the model's parameter count.