Deep attention models for visual recognition and completion

Abstract

The core of human vision is undoubtedly the perception ability, processing and translating the input visuals into structured, symbolic representations that provide the basis for higher-level understanding (visual recognition). Interestingly, this primary function is also designed to be robust on external interruptions. For example, humans can perceive and extract key information from the input, even if they are blurred, occluded, or damaged. We have a strong reasoning ability to examine the input visual signals against the corruptions (visual completion). Indeed, visual recognition and completion are crucial for understanding and interacting with our dynamic visual world. In this thesis, we first explore the effective ways to endow the machine with these two important human abilities and identify the key components to achieve this goal. Moreover, we investigate the synergy of bridging these two seemingly independent fields to further empower the machine vision ability and note the initial signals in this new direction.