Research On Algorithm Of Image Semantic Segmentation Based On Deep Learning

Image semantic segmentation is a pixel-level classification task.The goal of it is to determine the position and category of objects in the image accurately.In the semantic segmentation methods based on deep learning,convolutional neural network(CNN)is usually used to extract the image features and then have a classification by using classifiers.Finally,the classification results of the classifiers are modified by conditional random field(CRF)to improve the accuracy of pixel labeling.The problems to be solved in this method are:(1)how to extract better image features.(2)how to improve the efficiency of semantic segmentation.(3)how to use the information of objects' spatial position and color to improve the accuracy of pixel labeling.In response to the problems,this paper utilize CNN to construct the hierarchical features,the region-based features and accomplish the image semantic segmentation with superpixel and fully connected CRF.Details are as follows:Firstly,this paper propose a semantic segmentation method based on CNN hierarchical features.The convolution outputs of multiple layers in the CNN are upsampled to a uniform scale and combined to construct the hierarchical features.This hierarchical features integrate the structural information contained in the shallow layer of the network and the semantic information of the objects contained in the deep layer of the network,and therefore have strong expressive power.The superpixel segmentation algorithm can segment the image into multiple superpixels,train the classifier using the hierarchical features of the superpixels,and then map the classification results back to the pixels.Finally,a fully connected CRF including unary potential and pairwise potential is constructed,and the classification result of the pixels is smoothed by solving the corresponding energy function to improve the regional consistency and continuity of the pixel labeling.Experimental results on the public dataset show that the use of hierarchical features has significantly improved both average pixel accuracy and average class accuracy over single layer's features.Secondly,this paper propose a semantic segmentation method based on feature fusion and classifier fusion.Using the region-based neural network to extract the region-based features.This region-based features contain more details of local objects.The fusion features can be obtained with the fusion of hierarchical features containing more global information and region-based features containing more local information,this kind of fusion features can be better used for semantic segmentation.Using multiple different classifiers to classify the superpixels,the results of the classifier are weighted linearly to obtain the fusion result,which makes up for the limited classification ability of the single classifier.Finally,the smothness constraint of the fully connected CRF significantly improves the accuracy of pixel labeling.Experimental results on public datasets show that the feature fusion and classifier fusion can improve the average pixel accuracy and average class accuracy of image semantic segmentation.