Convolutional Neural Networks For Semantic Segmentation

Semantic segmentation is a kind of pixel level segmentation and classification of different objects in images,It is important for image processing and is widely used in different fields.A good algorithm of semantic segmentation can greatly promote the related algorithm researches.Deep convolutional neural networks has achieved remarkable results in machine vision in recent years,in the image classification task,it has made rapid development,repeatedly refreshed recordset,and surpassed the traditional algorithm of image classification task,and then transferred to the task of semantic segmentation.Deep learning is more advantageous than traditional methods in semantic segmentation.Traditional methods of semantic segmentation only concern at their specific characteristics,such as gray level,grain,gradient and so on.It has great limitations.Semantic segmentation based on deep learning can extract many kinds of feature of the target,and then make a dense prediction.The extraction of target feature is based on the parameters that have been learned,and it does not need to analyze the features of the target.Therefore,the application of deep learning in semantic segmentation has a strong realistic meaning.Aiming at the task of semantic segmentation of dense prediction,this paper proposes an improved method based on VGGNet in which the shallow information is fused with the deeper feature map,and feature extraction and fusion are carried out by using paralleling convolution with different sampling rates.This method is more effective in extracting features and context information from different layers,then improves the accuracy of semantic segmentation.This paper also improves the accuracy of semantic segmentation by optimizing the edge of the images with a fully connected conditional random fields.This paper has reached 71.2% mIOU in semantic segmentation task of test set of PASCAL VOC2012,and the results is superior to the previous main classical methods based on VGGNet.