Novel approaches to principal component analysis of image data based on feature partitioning framework

We present a feature partitioning framework for principal component analysis (PCA) on image data. Using this framework, we propose two novel methods, sub-image principal component analysis (SIMPCA) and flexible image principal component analysis (FLPCA). We prove the computational superiority of the approaches and also demonstrate improved performance through experimentation on standard face databases and a palmprint database. The proposed methods show a significantly superior performance as compared to conventional and improved implementations of PCA on images. It is seen that improvement in performance is in terms of both computational time and recognition rate. Experimentation shows that the novel partitioning approaches are in a different class of approaches. The success of proposed approaches may be attributed to the localization effect derived from partitioning. The proposed methods use a more appropriate matrix representation of the image data.