For an object classification system, the most critical obstacles towardsreal-world applications are often caused by large intra-class variability,arising from different lightings, occlusion and corruption, in limited samplesets. Most methods in the literature would fail when the training samples areheavily occluded, corrupted or have significant illumination or viewpointvariations. Besides, most of the existing methods and especially deeplearning-based methods, need large training sets to achieve a satisfactoryrecognition performance. Although using the pre-trained network on a genericlarge-scale dataset and fine-tune it to the small-sized target dataset is awidely used technique, this would not help when the content of base and targetdatasets are very different. To address these issues, we propose a jointprojection and low-rank dictionary learning method using dual graph constraints(JP-LRDL). The proposed joint learning method would enable us to learn thefeatures on top of which dictionaries can be better learned, from the data withlarge intra-class variability. Specifically, a structured class-specificdictionary is learned and the discrimination is further improved by imposing agraph constraint on the coding coefficients, that maximizes the intra-classcompactness and inter-class separability. We also enforce low-rank andstructural incoherence constraints on sub-dictionaries to make them morecompact and robust to variations and outliers and reduce the redundancy amongthem, respectively. To preserve the intrinsic structure of data and penalizeunfavourable relationship among training samples simultaneously, we introduce aprojection graph into the framework, which significantly enhances thediscriminative ability of the projection matrix and makes the method robust tosmall-sized and high-dimensional datasets.
展开▼
