A Semantic Segmentation Method Based On Improved U-net Network

Abstract

Semantic segmentation is the most important branch of image processing. Semantic segmentation based on deep learning occupies the mainstream method. CNN, FCN, U-NET and DEEPLAB series have achieved good results in the field of image semantic segmentation. The improvement methods of each network also emerge in an endless stream, but the emphasis of improving the network focuses on the feature transfer and feature fusion between network layers in each method. In this paper, based on u-net, we use deep convolution, residual autonomic force and dendrite network to improve the network model. Deep convolution can reduce the amount of calculation, deepen the depth of network, extract more abundant image information, and transmit information between network layers in a more refined way. Dendritic networks are different from ordinary neural networks. Ordinary neural networks process information as models of nerve cell bodies. The role of neural dendrites is also important in the process of information processing. Dendritic network is used as a model of neural dendrites to improve the ability of information processing. The improved network in this paper was tested on the VOC2012 dataset, and the experimental results showed that PA, mAP and IoU were 93.36%, 92.28% and 68.26% respectively. The experimental results showed that the improved network in this paper had improved in all indicators.