Semantic Segmentation Based On Convolutional Neural Network

With the explosion of image data and the continuous development of deep learning,computer vision has attracted the attention of many researchers.Under the impetus of the convolutional neural network,the accuracy of image classification,object detection and semantic segmentation tasks is improved steadily.The goal of semantic segmentation is partitioning image into different areas and also classifying each pixel,the traditional methods to solve the problem are extracting features on proposals artificially and then training the classifier.Semantic segmentation based on convolutional neural network is to train the net end-to-end directly.The automatic feature extraction based on convolutional neural network is more efficient and effective than the previous methods.After a full investigation,the new network is put forward.The appropriate strategy is added and the structure of the net is adjusted based on the classical convolution neural network model.Meanwhile,The proposed methods are compared on benchmark datasets.Specific research work and innovation points are as follows:Firstly,this paper introduces the basic principle of convolution neural network and the tool named caffe which is used to set up a net.Then the two kinds of semantic segmentation based on convolutional neural network are explained in detail,one is on the basis of the classification of image block,the other is based on fully convolutional neural network.The two classic methods are evaluated on the PASCAL VOC 2012 datasets.This chapter provides the theoretical and practical basis for follow-up study.Secondly,this paper proposes a semantic segmentation algorithm based on multi-scale pooling strategy.The detection algorithm based on multi-scale pooling strategy is first introduced.Then the strategy is utilized to the semantic segmentation algorithm for better results.It solves the problem that the target can not be segmented and recognized completely by using multi-scale input and multi-scale pooling.The multi-scale strategy changes the receptive field of the network and helps improve the accuracy.The proposed semantic segmentation method is evaluated on two datasets and experimental results support the proposed method.Thirdly,This paper proposes a CRF_integrated deconvolution network method for semantic segmentation.The great advantage of deconvolution network is that it distinguishes between different sizes,and the smaller or bigger objects will rarely be fragmented or mislabeled.Conditional Random Fields is interpreted as Recurrent Neural Networks,which is then integrated into deconvolution.The deconvolution and CRF are trained jointly,averting the post-processing with CRF to optimize the edges of the objects.Because of the joint training and the more accurate unary potentials,the parameters of CRF are more robust.The experimental results of the proposed method show good performance on benchmark datasets.