MFA-Deeplabv3+: an improved lightweight semantic segmentation algorithm based on Deeplabv3+

The widely used DeepLabv3 + network for semantic segmentation suffers from issues such as excessive parameters, imprecise segmentation of edge targets, and difficulty in deployment on mobile devices. This paper proposes an improved lightweight semantic segmentation algorithm, MFA-DeepLabv3+, based on DeepLabv3+. The proposed MFA-DeepLabv3 + employs MobileNetV2 as the backbone network to reduce model parameters and computational costs. A multi-feature fusion (MFF) structure is designed to expand the receptive field, while optimizing the atrous rates in the Atrous Spatial Pyramid Pooling (ASPP)module to enhance spatial information extraction and mitigate critical information loss. Additionally, a Global Pyramid Attention (GPA) module is introduced to improve context-aware feature capture, enabling the network to focus on key image regions through adaptive information filtering. Experimental results demonstrate that on the PASCAL VOC 2012_aug dataset, the model achieves improvements of 1.64% in mIou and 1.21% in FWIoU. On the Cityscapes dataset, it attains gains of 2.75% in mIou and 1.42% in FWIoU, respectively reducing the parameter count by 83%.