The Image Semantic Segmentation Algorithm Based On Deep Neural Networks

In the field of computer vision,image semantic segmentation has always been a hot topic.As an important part of image cognition,the accuracy of semantic segmentation directly affects the results of image understanding and scene perception.Since it can not only segment stuffs,but also provide semantic information,it plays a crucial role in various fields,such as the scene understanding in the area of automatic driving or human segmentation in the area of entertainment.In recent years,with the development of deep learning algorithm and hardware equipment,breakthroughs have been made in images semantic segmentation.Different from other methods,semantic segmentation results can be achieved end-to-end with the help of deep neural network and its accuracy is far higher than other methods.In this paper,we also study the image semantic segmentation based on the deep neural network,including fully-supervised image semantic segmentation and semi-supervised image semantic segmentation.Semi-supervised image semantic segmentation is an extension of the task of fully-supervised image semantic segmentation,which aims to solve the problem of large requirement of labeling pixel-wise semantic information for fully supervision.In semi-supervised learning,unlabeled data is utilized to improve the accuracy and generalization of a pretrained network.Firstly,in the research of fully-supervised image semantic segmentation,we propose a Multi-Scale Recurrent Network(MSR-net)from the perspective of network structure and loss function.In order to extract classification information and location information at the same time,we propose the Spatial Pyramid Recurrent module(SPR)which based on spatial pyramid structure.Meanwhile,we find that there is relationship among features on different scales which can improve the ability to extract classification information and location information.In order to extract this relationship,we add Recurrent Neural Network(RNN)in SPR.In order to eliminate the inference of using image classification structure as backbone,we propose the Feature Fusion Module(FF)which fuses low-dimensional features from backbone and the high-dimensional features based on attention mechanism.Moreover,we propose the Semantic Classification Loss(SC Loss)which takes the classification of stuff as supervision.It can r educe the impact the effect of different size of stuffs and push the network to extract classification information.Secondly,in the research of semi-supervised image semantic segmentation,we propose a Semi-Supervised Semantic Segmentation Method based on Uncertainty and Conditional Random Field(CRF)based on pseudo-groundtruth algorithm.In order to eliminate the error in pseudo-groundtruth,we utilize the uncertainty to measure the level of accuracy.In this paper,advanced Bayes Cross Entropy Loss and advanced CRF is used to get uncertainty and refine the pseudo-groundtruth respectively.Moreover,according to the meaning of uncertainty,we propose two criterions of uncertainty and propose the Uncertainty Loss from them which is taken to implement semi-supervised learning.Finally,we evaluate the proposed MSR-net in Pascal VOC dataset and Semi-Supervised Semantic Segmentation Method based on Uncertainty and CRF respectively in human segmentation dataset.These experiments show the validity of method of MSR-net and Semi-Supervised learning.More importance,our results achieve the state-of-art performance in Pascal VOC dataset when we add our module into Deeplab v3+.