Improved DeepLabv3+ for Semantic Segmentation via Multi-Scale Context Fusion and Feature Attention

Abstract

Aiming at the problems of small object edge contour detail loss and insufficient feature representation capability in semantic segmentation tasks, the paper proposes an improved DeepLabv3+-based semantic segmentation algorithm, termed RDSC-DeepLabv3+, which integrates multi-scale contextual enhancement with feature attention optimization. Based on the original DeepLabv3+ framework, the model architecture is improved from multiple aspects, including backbone network selection, multi-scale contextual modeling, and feature attention enhancement: (1) ResNet-50 is adopted as the backbone network to reduce the number of model parameters and computational complexity while maintaining strong feature representation capability;(2) a multi-scale feature fusion module that combines DenseASPP with Strip Pooling (SP) is introduced to enhance the model’s ability to capture contextual information of objects at different scales;(3) the CBAM attention mechanism is incorporated into shallow features to adaptively enhance discriminative region features while suppressing redundant background information.Experimental results on the PASCAL VOC 2007 dataset demonstrate that, compared with the baseline model, the proposed method achieves improvements of 1.86%, 1.78%, and 1.49% in ACC, mPA, and mIoU, respectively. These results verify the comprehensive advantages of RDSC-DeepLabv3+ in terms of segmentation accuracy and computational efficiency, while also exhibiting stronger boundary delineation capability and robustness in complex scenes.