Semantic Image Segmentation Algorithm Based On Deep Learning

Semantic segmentation has been a challenging task in computer vision,and plays a central role in image understanding.Semantic segmentation combines the image segmentation and image understanding to segment objects accurately and give a semantic label to each pixel in the image.Semantic segmentation can obtain the expression quickly and effectively in video or image content,and provide useful information for Internet users.In recent years,deep learning has been a hot research area in machine learning.It has strong capacity of features learning expression.Convolutional neural network can extract image information from pixel-level data to abstract semantic level,which makes the deep learning model has outstanding advantages in terms of the extraction of image context information.Deep learning is also a new way to solve the problem of image semantic segmentation.In this paper,inspired by the popular deep learning model,we use fully convolutional networks,and combine with the high order conditional random field with target detection information,to achieve the semantic segmentation.(1)We summarize the current research situation of semantic segmentation,the basic concept and principle of conditional random fields(CRF)models.Furthermore,we point out the existing advantages and shortcomings in conditional random fields,and lead to higher order conditional random field model,which can effectively improve the accuracy of image segmentation.(2)We discuss the theory and network structure of convolutional neural network,and introduce fully convolutional networks for semantic segmentation.We apply the technology to automatic segmentation of the prostate MRI image.The experiments show that this popular network model is better than the traditional algorithm.(3)The pairwise conditional random field could cause boundary segmentation errors and recognition errors in semantic segmentation.To solve of these problems,we propose a higher-order CRF model combined with detection information,and use fully convolutional networks to extract the image feature.The training of the high-order CRF model and the convolution network are unified in a system framework to improve the performance of the model.The experimental results on MSRC-21 and PASCAL VOC2011 database show that the performance of our proposed segmentation algorithm field model is better than other algorithms.