Research On The Application Of Sparse Representation In Image Classification Problem

As the Internet and the intelligent terminals have become more and more popular, the data of image is also growing rapidly. But because of the lack of annotation for those images, it is inconvenient for the image storage and search. How to classify large amount of images effectively has become a popular topic in the field of image processing.Sparse representation applied in image classification is an important branch in image processing. Sparse representation with the simple model quickly attracted the attention of many scholars. Through sparse representation coefficients vector, reconstruction of sample in each subclass dictionary and record the residuals of test sample minus all reconstructed sample, find the smallest residuals corresponding to the sub dictionary, for the predicted labels of the test sample, the image classification method is intuitive. For a good dictionary samples, through coefficient can intuitive observation a test sample is associated with which classes.This thesis based on the application of sparse representation in image classification problem, the main work and innovation are as follows:(1)An image classification algorithm based on sparse representation of reconstructed samples (ICSRIR) is proposed, the main idea of the algorithm is classified by producing a reconstruction the middle sample. In view of the existing in the existing sparse representation based image classification is not all information can used to classification problems, put forward a kind of test samples are projected to the dictionary, dropping some training samples are not represented in the information, and after reconstruction do the sample classification, to effectively solve the training samples cannot represents some information. In order to achieve better results, all of the training samples and test samples are extracted local features through the Gabor filter. The experimental results show that the method has certain effectiveness.(2) A method of face image classification combined projective dictionary pair learning sparse representation with image reconstruction (DPL_ICSRIR) is proposed.For image classification and recognition, it is essential to obtain a such dictionary that possess the good property of distinguishing. In view of the existing image classification based on sparse representation has high complexity to get discriminating dictionary and the test sample cannot be represented in all dictionary problems, a sparse representation for image classification is proposed, which is based on the sparse representation of dictionary learning and sample reconstruction. In the training process for the training samples and test samples, using Gabor filter from different directions, different scales as the original input training dictionary firstly, and then use the analytical dictionary to obtain comprehensive dictionary has the ability to distinguish between, then test samples throw away some information which dictionary without. Reconstruction samples based on the comprehensive dictionary for classification, reducing the impact of information cannot be said on the dictionary, improve the accuracy of classification. Experimental results show in three representative face datasets, DPL_ICSRIR method can well classify face image.