Image Denoising And Face Recognition Algorithm Based On Sparse Representation

Sparse representation is one of the hottest topics in the field of signal processing in recent years. In simple terms, sparse representation is a process of decomposition of the original signal which represent the input signal as linear approximation of the dictionary by means of a prior dictionary(also known as over-complete basis).Sparse representation of high dimensional data is a hot topic of research in machine learning and computer vision in recent years. It is generally assumed that the natural image itself is a sparse signal, when the input signal expressed linearly by a group of over-complete basis, the expansion coefficients can obtain a good approximation of the original input signal under certain sparsity conditions. This method has achieved a great success in image denoising, image blur and image high resolution etc. With the deepening of research, people found that although the optimal model coefficients is expressed by the reconstruction of a signal from the angle of design, but its expression results in the identification of model have good performance.Therefore many of the current current classification system often tend to choose the best sparse representation as its key module.In this paper, some improvements have been made in the existing sparse representation model including image denoising and face recognition. They are GNCSR model used in image denoising and TSRC model used in face recognition.Due to the single structure dictionary, the result in the original image restoration model is instability, that is to say we may get different results. So, we try to change the method to build dictionary which would obtain stable and better results. In GNCSR, we firstly classify the test image, and then code each class using Analysis K-SVD sub dictionary. For each given block, GNCSR method firstly calculate distance between it and the mean class to determine if it belongs to one class, and then use the corresponding sub dictionary to encode them, which does avoid the single dictionary.Since conventional sparse representation overrely sparsity, to a certain extent, its recognition rate is increased with the increase of the capacity of training samples. However, when the sample size is small or the data correlation is high, the SRC face recognition method can not get good recognition performance. In order to fully consider the sparsity and correlation of the sample data, this paper puts forward the concept of trace norm regularization term, and it is proved that the value of the term is between 1l-norm and 2l-norm. Adding the term to SRC face recognition model, we propose TSRC face recognition model. It can be seen that TSRC method can get better recognition results than other methods through experiments in different face databases.