Optimization of the DeepLabv3+ Segmentation Network by Integrating Multi-Scale Spatial Information

Existing algorithms fail to fully utilize the rich semantic and spatial information contained in images, leading to inaccurate pixel segmentation across different categories and severe loss of detail. We propose a semantic segmentation network that focuses on multi-scale spatial information (Focusing on Multi-Scale Spatial Information for Semantic Segmentation Networks, FMSI-DeepLab). Based on the DeepLabv3+ framework, the network is improved in two main parts. In the encoder, deformable convolutions are combined with the Global Grouped Coordinate Attention (GGCA) mechanism to reconstruct the Atrous Spatial Pyramid Pooling (ASPP) module, enhancing the model's ability to capture global information across both height and width spatial dimensions, thereby enabling efficient multi-scale feature extraction. In the decoder part, "interest flow" processing is added to the low-level features, enabling them to have global connectivity at the low-level stage. Subsequently, the Multi-Scale Channel Spatial Enhanced Attention (MSEA) module is introduced to further enhance the model's focus on the edge information of the low-level features extracted by the backbone network, thereby strengthening the model's emphasis on details. Compared to the original DeepLabv3+ semantic segmentation model, the model achieves a 2.62% improvement in average intersection-over-union (mIoU) on the VOC2012 dataset, addressing issues of inaccurate image segmentation and severe loss of details.