Combination Of Deep Metric Learning And Conditional Random Field

One of the most important research fields in computer vision is image classification,which is the foundation of many other advanced visual tasks,such as face recognition and image retrieval.In recent years,one of the hot topics in image classification is hyperspectral image classification.Because the hyperspectral image contains high spectral resolution,making the ability to distinguish and identify the ground objects improved greatly,so it plays an important role in agriculture,ocean,transportation,and other aspects.Traditional machine learning methods rely on artificially designed shallow features for completing the classification,which takes a lot of time and labor,and these methods perform unsatisfactorily when they encounter high-dimensional data.However,deep learning greatly improves the performance of traditional classification methods because of its ability to learn abstract features autonomously.Based on deep learning,deep metric learning provides a similarity measurement for data,making the generated features more discriminative,and thus better for improving the classification accuracy.Besides,the conditional random field of the probability graph model has also got good achievements for classification.With the full use of context information,it has achieved better results than the general classification methods which depend on pixel features only.To this end,a method combining deep metric learning and conditional random field is studied in this paper,and also studies the performance of the proposed method for classifying hyperspectral images.The main work in this paper is described as follow:1.A deep metric learning model for feature extraction is proposed.The deep metric learning model uses a neural network framework.On the basis of the original Softmax Loss,Center Loss is introduced to jointly train the network in a supervised manner.After training,the network can extract the features with the characteristics of compactness within a class and separation between classes from the input data.The experimental results show that the extracted features by the proposed method are more discriminative and the classification performance is better.2.A classification method combining deep metric learning and conditional random field is proposed.To further improve the classification performance,when considering the rich context information in the image,we employ conditional random field as the post-processing classification step.The deep metric learning model is used to extract features,and the conditional random field model is used to complete the final classification according to the extracted features.The experiment results also demonstrate that the classification accuracy can be further improved after fusing the spatial context information.3.The ConvCRF algorithm for fast inference of the conditional random field model is employed.Because the conditional random field model contains a huge number of edges,making it possible to direct inference.To this end,we use the ConvCRF algorithm which based on a mean-field approximation method to perform the inference.Experiments show that the proposed method is more efficient than other deep learning models in terms of computing cost.