Enhanced Semantic Segmentation with Hierarchical Upsampling and CBAM Attention Mechanism

Deeplabv3+ is a commonly used image semantic segmentation network designed to extract strong semantic information. The traditional Deeplabv3+ network, due to stacking all features together, struggles to focus on important features and the spatial relationships between different regions. Furthermore, its decoder performs upsampling by a factor of four twice, leading to the loss of many details. To address these issues, we propose a hierarchical upsampling structure combined with Inverted residual and introduce the CBAM mixed attention mechanism into the ASPP module. The hierarchical upsampling structure preserves more detail, and CBAM focuses on more important features. Experimental results show that the mIoU value of this algorithm is increased by 4.14% and the accuracy is improved by 0.88% compared to the original algorithm.