General image segmentation by Deeper Residual U-Net

With the development of deep learning, using convolutional neural networks for semantic segmentation has received a large amount of attention. Numerous convolutional neural networks architecture has been proposed. In biomedical image processing, U-Net has achieved great remarkable achievement. However, due to the feeble convolution operations for extracting complex image information, the U-Net presents a poor performance in general semantic segmentation. Therefore, in this paper, we propose a new neural network framework, called 'Deeper Residual U-Net' for general image semantic segmentation. In our method, we apply ResNet101 for extracting features and use a double features fusion mechanism compared to U-net. In the first time, the Deeper Residual U-Net up sample each stage features and fuses them with features of the previous layer one by one, which make low-level features contain more abstract information. In the second time, it upsamples all fused features of different stages to the same size and combines them to predict. We test our network in Pascal VOC 2012 dataset and get mean accuracy 80.9, mIoU accuracy 74.3, which already available for general image segmentation.