Deep learning based visual recognition robust against background clutters

Abstract

Traditional obstacles that make the visual recognition problem difficult are the occlusion, within-class diversity, inter-class similarity, geometric/photometric transformations, and background clutters. This dissertation focuses on the background clutters. The recent evolution of deep learning has significantly suppressed the negative effect of backgrounds by learning common visual features of objects from big data that cover diverse backgrounds. However, a deep learning based recognizer still lacks an explicit design to solve this problem but recognizes objects by detecting foreground features common to objects. Therefore, the ability to cope with the background is limited to the background distribution of the training data. In this dissertation, we define problems of background clutters in various visual recognition tasks and presents solutions based on deep learning. The background can be an obstacle to visual recognition, but it can also be an important clue to determine the object existence. We simultaneously take these two aspects of backgrounds into consideration to improve the performance of object classification and detection. Once a target object is classified or detected, the background always acts as an obstacle for another recognizer to analyze attributes of the object. To solve this problem, we learn a normalized object appearance invariant to the backgrounds. When we apply these recognizers to the real world, they often fail to recognize objects with new backgrounds out of the training background distribution. To maintain the recognition performance, we propose an active background adaptation method that enables a recognizer to adapt quickly to the new backgrounds.