Image Semantic Space Segmentation Based on Cascaded Feature Fusion and Asymmetric Convolution Module

Abstract

With the rapid development of deep convolutional neural networks, the results of image semantic segmentation are remarkable, and the segmentation effect is greatly improved. The pooling layer of the convolutional neural network will reduce the resolution of the feature map, which makes the convolutional neural network lose a lot of spatial information while extracting semantic features. How to integrate semantic features with semantic information and spatial information will become an important factor to improve the performance of semantic segmentation. Firstly, this paper improves the global attention upsampling module and uses the improved global attention upsampling module to form a multiscale global attention up-mining module in a new connection way. The upsampling module of multiscale attention establishes the relationship between high-level features and lower-level features at a longer distance. Compared with PANet, the method proposed in this paper deepens the close relationship between semantic information and spatial information. Experiments show that the segmentation effect of the feature fusion method based on cascade is better than that of the feature fusion method based on weight. The segmentation effect of the two fusion methods is improved by 8.3% and 5.7% compared with the PANet on the PASCAL VOC 2012 dataset and by 4.5% and 3.6% on the Cityscapes dataset, respectively. The research results of this paper make the high-level semantic information and shallow feature information cooperate to improve the segmentation effect.