High precision semantic segmentation based on multi-level feature fusion

Abstract

In order to solve the problems of fuzzy edge segmentation and imprecise segmentation of small objects in image semantic segmentation, a high-precision semantic segmentation method was proposed. The method used MobileNetV3 network to extract multi-level shallow contour features and deep semantic features, and then fused the multi-level shallow contour feature information with the deep semantic feature information by Pyramid pooling module and up-sampling operation in PSP-Net model, so as to realize the high-precision semantic segmentation with multi-level feature fusion. The experimental results on the Nyu-V2 dataset show that the proposed algorithm can improve the description ability of small target features. The generalization of the proposed algorithm is further verified on the Pascal-VOC2012 dataset. Experimental comparison with three mainstream methods shows that the segmentation accuracy of the proposed algorithm increases by 2.1% compared with Deeplabv3+, by 5.1% PSP-Net and by 10.9% SEG-Net.