Image classification is a tough research topic in the field of computer vision and pattern recognition, which firstly extracts information from images via image processing technique and then automatically performs classification by machine learning methods. This study focuses on local features and dictionary learning based approaches for image c1 assification. Specifically, the thesis proposes local features and dictionary learning based methods and tests with LFW face and HEp-2 cell databases.In terms of features, this paper not only introduces the classical SIFT features, also introduces a novel feature called Gabor Ternary Pattern(GTP) which is based on multi-scale Canny detector. Experiments show that both features are robust to rotation and scale variance, and Canny based affine invariant detector generates more key points than SIFT.To overcome the problems of occlusion or pose variation in face recognition, an alignment free approach based on joint sparse representation is proposed. We conduct experiments on the publicly available LFW face database and get a promising result. The general framework is also applicable to HEp-2 cell classification. Experiments show that the GTP performs better than SIFT in cell classification.Consider that the dictionary of joint sparse representation is constructed without learning, this paper proposes to use clustering-technique used in BOW model for dictionary learning. Benefit from dictionary learning, the new approach achieves a higher accuracy than the joint sparse representation framework. |