Image Semantic Segmentation Based on Depth Parallel Convolutional Networks

Abstract

Fully convolutional network can effectively learn the task of centrally predicting pixels, such as semantic segmentation. However, FCN has the problems of spatial invariance and non-real time, which limits its performance and application. We propose an improved structure that preserves the basic shape of the FCN while extracting high-level features in the input section and introduces a separate residual learning module. On the other hand, we also introduce sub-networks that exist in parallel and reduce the size of the original FCN structure by proportional random discarding. Experiments show that the proposed method, which increases the residual network and parallel FCN and random discarding has better evaluation index than the traditional FCN method, and can further reduce the computational cost of the system with ensuring performance.