Research On Key Technologies Of Structured Sparse Coding In Image Processing And Recognition

With the development and maturation of imaging techniques,visual data have become more and more popular in various fields and are increasing at an explosive rate.Among various types of visual data,images are the most common one.How to discover useful information and knowledges by processing and analyzing images is one of the most important topics in the age of intelligence.One key step in image processing and analysis is to find effective representations of images in which the inherent structures and specific patterns of images can be revealed.In recent years,sparse coding has become a promising way for representing images,which finds the sparse representation of data under some dictionaries and obtains the underlying patterns.Though there have been a lot of studies on sparse coding,it is still a hot issue how to construct structured sparse coding models for different types of visual tasks,such that induces useful patterns in sparse codes to improve both the effectiveness and efficiency.This paper studies the key technologies of structured sparse coding for image recognition,image sequence processing and image separation,which includes:1.For the recognition task,a novel label consistency sparse coding method is proposed.By introducing sparse label consistency constraints,an effective explicit structured sparse coding method is achieved,which overcomes the defects of traditional label consistency methods and greatly improves the discriminability of sparse codes.2.For the recognition task,an implicit structured sparse coding method is proposed based on the multiple linear classifier integration strategy.The method constructs a discriminative loss term based on the integration of multiple linear predictors in the coding space.The method is capable of characterizing the complex non-linear classification boundary in coding space effectively and meanwhile guaranteeing the ease of the solution.3.In order to further improve the ability of sparse coding for processing the nonlinearity of data and reducing the dependence on training data,a joint computational framework for sparse coding and decision forest is proposed.The sparse coding is used to improve the representation ability of decision trees.Meanwhile,the feedback of the decision trees is used to improve the discriminativeness of sparse codes.Finally,an efficient sparse coding based decision forest is generated for recognition task.4.By constructing an orthogonal tensor structured sparse coding model,a fast sparse coding and dictionary learning method for image sequence data is proposed,which overcomes the disadvantages of traditional sparse coding methods such as high computational complexity and low scalability when processing high-dimensional data.Based on the proposed dictionary learning method,a feature descriptor with high adaptability,strong discriminability and high scalability is designed for dynamic texture data.5.In order to remove the reflection from images,an image decomposition model is constructed based on structured sparse coding.Through the sparse regularization on the clear image as well as the analysis on the unique structure of the reflection layer,a weighted structured sparse coding method is proposed,which can effectively separate the reflection layer from the input image.The proposed methods were applied to multiple visual tasks with performance evaluation.Experimental results have demonstrated the effectiveness of the proposed methods.The results of this paper have significant importance for the development of sparse coding and dictionary learning,and they also provide new insights and ideas to structured sparse representation from the aspects of theories,models and algorithms.