Image Segmentation Based On Convolution Neural Network And Conditional Random Field

Image segmentation is a basic and difficult problem in low-level image processing.As the development of the infrared imaging technology,image segmentation technology is playing a more and more important role in the fields of medicine,satellite and computer vision.As a multidisciplinary subject,the image segmentation has been faced with the challenges of feature extraction and data annotation.And the emergence of Convolutional Neural Networks(CNNs)alleviate these problems.This paper mainly deals with the image segmentation by combining CNNs with conditional random field,which is of great significance in theory and application.The main research work and innovation of this paper include the following:First of all,the theories of image segmentation related to deep learning are studied.By summarizing and contrasting the traditional theory and the latest theory of deep learning,we propose an image segmentation learning framework based on deep learning and probabilistic graphical model.Then,a supervised model based on FCN-FCRF combining the high-level layer with the low-level layer is proposed.To extract features,we transform the fully connected layer into the convolution layer to construct the Fully Convolution Network and combine the high level layer with abstract features and the low layer with fine features.With the features as input,the CRF can complete the tasks of context modeling,edge precise positioning and classification.The model is trained with end-to-end by sensitivity transferring the sensitivity of neuron.The experiments show that when compared with others',our model has achieved better results,which demonstrated that our model extract image feature more comprehensive,position edges more accurately and train model parameters more consistency.Finally,a weakly/semi-supervised model based on dense sampling FCN model is proposed.Because pixel-level annotation data are time-consuming and expensive,we use weak and weak-semi supervised data for image segmentation modeling.We build a dense sampling FCN,and design three kinds of segmentation model,including weak supervised model for image-level labels based on the EM algorithm,weak supervised model for bounding boxes based on the GrubCut algorithm and weak-semi supervised model for mixed image-level labels and pixel-level annotation labels based on the EM algorithm.The experimental results indicate that our weakly supervised model has more accurate results than other similar models,and our weak-semi supervised model can be close to the results of full supervised model.