Real-time efficient semantic segmentation network based on improved ASPP and parallel fusion module in complex scenes

Semantic segmentation can help the perception link to better build an understanding of complex scenes, and can assist the unmanned system to better perceive the scene content. To address the problem of detailed information loss and segmentation edge blur in the semantic segmentation task for complex scenes, we propose a modified version of Deeplabv3+ based on the improved ASPP and fusion module. Firstly, we propose an RA-ASPP module combining residual network and asymmetric atrous convolution block (AACB), which further enriches the scale of feature extraction and achieves denser multi-scale feature extraction. It significantly enhances the representation power of the network. Then, we propose a parallel fusion module named convolution combine with bottleneck block (CBB), which combines 1 × 1 convolution and bottleneck block to reduce the information loss in the whole network transmission process. We perform ablation experiments on the PASCAL VOC2012 dataset. When the backbone is Xception, the Mean Intersection over Union (MIoU) of Ours1 is 79.78 % . At the cost of 1.72 frames per second (FPS), its MIoU is 2.81 % faster than Deeplabv3+. The proposed modules significantly improve the accuracy in semantic segmentation and achieve segmentation results comparable to state-of-the-art algorithms. When MobileNetV2 is the backbone, Ours2 achieves 37.54FPS and a MIoU of 73.32 % , which ensures a balance between real-time segmentation speed and accuracy. In summary, our proposed modified module improves the segmentation performance of Deeplabv3+, and the different backbones also provide additional options for semantic segmentation tasks in complex scenes.