CMANet: Convolutional Multi-channel Fusion Attention Mechanism Network for Semantic Image Segmentation

Semantic segmentation is a significant task in the field of computer vision (CV). The DeepLabv3+ network has achieved advanced performance in semantic segmentation tasks, but also faces problems such as blurry segmentation boundaries, loss of small target segmentation, and incomplete segmentation of large targets. To address these issues, a Convolutional Multi-channel Fusion Attention Mechanism Network (CMANet) has been devised. Firstly, the backbone network extracts the feature maps of three different channels in the coding stage, namely low-level, mid-level and high-level semantic information feature maps. The input image undergoes varying degrees of feature extraction operations, followed by multi-channel feature fusion. Secondly, the designed multi-scale fusion atrous space pyramid pooling module adopts multi-layer atrous convolution with different expansion rates for feature extraction, and performs multi-stage feature fusion at the same time to avoid the loss of initial feature information when extracting semantic features, resulting in segmentation targets with voids. According to the experimental results of CMANet in the dataset PASCAL VOC 2012, the proposed model reaches 95.8\%, 83.5\%, 93.8\% and 92.3\% in the four evaluation indicators of Pixel\_Accuracy, mean Intersection-over-Union (mIoU), Precision and Recall, respectively, which are 2.5\%, 9.9\%, 7.5\% and 9.1\% higher than the DeepLabv3+ model, respectively.