Image Semantic Segmentation Algorithm Improved Based on DeepLabV3+

At present, the DeepLabV3+image semantic segmentation method is challenging to implement on embedded devices with limited processing power due to its complex computational requirements and memory requirements. When extracting image feature information, DeepLabV3+does not completely exploit multilevel information, which can lead to the loss of some details and reduce the accuracy of segmentation performance. Taking lightweight MobileNetV2 for the model backbone, this study suggests a better approach based on the DeepLabv3+ network. The Efficient Multi-Scale Attention (EMA) module is applied to high-level features to enable the network to obtain higher pixel-by-pixel attention to feature maps with high resolution. To capture long-range dependency relationships, the Strip Attention Module (SAM) takes the place of the global average pooling that is part of the ASPP module, and the Sampling-based Spatial Feature Calibration (SFC) module is used to fuse low-level and high-level features. Validation on the PASCAL VOC2012 dataset shows that the improved mIoU and mPA reach 72.35% and 83.14%, respectively. It can predict more refined semantic segmentation results, successfully lowering the model's complexity and raising segmentation accuracy.