Research On Semantic Segmentation Algorithm Based On Deep Features

Image semantic segmentation is an important research branch in the field of machine vision.Recently,image semantic segmentation based on deep neural networks has become the focus of researchers.But as the research goes deep,the segmentation algorithm based on deep neural network gradually reveals some problems such as missing the object in image,segmentation result is not elaborate enough.Aiming at the above problems,a lot of research has been done in this paper.In the survey,we found that the image segmentation based on conditional random fields performs well on the boundary segmentation of objects.The deep convolution neural network combined with attention module has the excellent expression ability of image semantic features and the excellent detection ability of the object.Based on the above understanding,the main contents of this paper are as follows:1,Combined with the advantages of full convolutional network and fully connected conditional random field,taking the two as the basic component of this paper's algorithm,we design a semantic segmentation algorithm framework which includes both global and semantic information of images.2.The attention module recalculated the relative importance of image features,which could highlight the area of the object in the image.According to the good expression ability of attention module to object characteristics in images,a deep neural network combined with attention module is designed for network topology.3.According to the Markov characteristics of the conditional random field,the image segmentation technique based on dense conditional random fields is segmented well in the details of the image,thus providing a way of thinking for the rough boundary problem of semantic segmentation.Based on the characteristics of the dense conditional random field model that can model the details of the image,the spatial context pixel level modeling method of the object is designed.4.Research on local information modeling and deep feature fusion technology of conditional random field.The probability score diagram obtained from the deep convolutional neural network contains the global semantic information.The dense conditional random fields can model the relationship between pixels based on semantic features and extract local details information.By applying the dense conditional random field on probability score map and original image,local information and global information can be effectively combined.Finally,the experiments of semantic segmentation and object region segmentation are carried out on different experimental data sets.The experimental results show that the method can accurately segment the boundary of the image and get a more complete target area,which effectively solves the problem of rough boundary segmentation using the depth feature alone.