Abstarct: With ever increasing digital images on personal computers and web servers, the image classification is considered more than ever. Accordingly in recent years various features and also classifiers is presented to solve this problem. Combining various features with the classifier baxsed on neural network is one of the suggestions that is presented in this thesis. Classification baxsed on sparse representation is another method that we will discuss here, and according to the novelty of this approach, we will further focus on it. In this method, within a large number of basis signals that are much greater than their dimensions, in general; we select the minimum number of basis to display a signal. Each basis signal is called an "atom" and set of these atoms is called a "dictionary". It’s very hard that select the best and lowest atom to represent a signal. But in recent years, researchers optimized both the speed and accuracy of this method. Thus, this method quickly found several applications in signal processing such as compression, image denoising, pattern recognition and authentication. Before solving a problem using sparse representation, we must consider two important issues. The first issue is finding an appropriate dictionary for the data and the second issue is to find an efficient algorithm to obtain a sparse representation of the signal. In this thesis, we first review a number of well-known algorithms in face recognition and then represent the issue of sparse classification, dictionary learning and sparse coding. Next, we propose a method baxsed on neural network and also will offer a method inspired by sparse representation. The simulation results show the promising performance of this methods. For example, One of the methods that are proposed in this thesis present recognition rate 98.13 % on the ORL databaxse.