This thesis examines a theoretical framework for forensic face recognition that is subject to a wide range of internal/external impact factors to achieve face recognition under different views, illuminations, resolutions, modalities, periods, when probe images are captured in surveillance environments without collaborations.
Within the scope of forensic photographic comparison techniques, forensic face recognition has the goal of identifying suspects from a large amount of gallery photos for the current probe image, which could be a sketch based on witness descriptions, a captured image from surveillance videos or street-shot photos from passersby. This thesis examines a theoretical framework for forensic face recognition that is subject to a wide range of internal/external impact factors to achieve face recognition; it explores the factor of view variance, which poses a challenge for forensic face recognition, and proposes a multi-view recognition that covers two settings in multi-view face recognition. The thesis addresses another challenge faced by analysts when they encounter already-corrupted data, and introduces a novel Deep Robust Encoder (DRE) through locality preserving low-rank dictionary to extract robust and differentiating features from corrupted data. Lastly, the dissertation proposes a one-shot generative model to build a more effective face recognizer. The dissertation is organized by chapter, with the first providing the introduction; chapter two develops a view-invariant framework based on both subspace and deep learning to address the test multi-view data with view information unknown, exploring multiple view-specific structures and one view-invariant structure; chapter three utilizes the knowledge of one domain to do face recognition on another domain, considering the missing modality problem, it uses a two-directional transfer learning framework to iteratively seek a domain-invariant feature space; chapter four develops a deep feature learning framework tin order to seek better feature representation for face recognition; chapter five addresses the need to solve one-shot face recognition where only one training sample is available for some persons in the training stage; and chapter six provides conclusions on what the author proposes in order to solve forensic face recognition in various challenges.