Face Recognition, a very challenging research area, is being studied for almost more than a decade to solve variety of problems associated with it e.g. PIE (pose, expression and illumination), occlusion, gesture, aging etc. Most of the time, these problems are considered in situations when images are captured from same sensors / cameras / modalities. The methods in this domain are termed as homogeneous face recognition. In reality face images are being captured from alternate modalities also e.g. near infrared (NIR), thermal, sketch, digital (high resolution), web-cam (low resolution) which further alleviates face recognition problem. So, matching faces from different modalities are categorized as heterogeneous face recognition (HFR). This dissertation has major contributions in heterogeneous face recognition as well as its homogeneous counterpart. The first contribution is related to multi-scale LBP, Sequential forward search and KCRC-RLS method. Multi scale approaches results in high dimensional feature vectors that increases computational cost of the proposed approach and overtraining problem. Sequential forward approach is adopted to analyze the effect of multi-scale. This study brings an interesting facts about the merging of features of individual scale that it results in significant reduction of the variance of recognition rates among individual scales. In second contribution, I extend the efficacy of PLDA to heterogeneous face recognition. Due to its probabilistic nature, information from different modalities can easily be combined and priors can be applied over the possible matching. To the best of author’s knowledge, this is first study that aims to apply PLDA for intermodality face recognition. The third contribution is about solving small sample size problem in HFR scenarios by using intensity based features. Bagging based TFA method is proposed to exhaustively test face databases in cross validation environment with leave one out strategy to report fair and comparable results. The fourth contribution is about the module which can identify the modality types is missing in face recognition pipeline. The identification of the modalities in heterogeneous face recognition is required to assist automation in HFR methods. The fifth contribution is an extension of PLDA used in my second contribuiton. Bagging based probabilistic linear discriminant analysis is proposed to tackle problem of biased results as it uses overlapping train and test sets. Histogram of gradient descriptors (HOG) are applied and recognition rates using this method outperform all the state-of-the-art methods with only HOG features.
|Publication status||Accepted/In press - 1 Nov 2015|