Multistrategy Learning for Computer Vision

Abstract : Current IU algorithms and systems lack the robustness to successfully process imagery acquired under real-world scenario. They do not provide the necessary consistency, reliability and predictability of results. Robust 3-D object recognition, in practical applications, remains one of the important but elusive goals of IU research. With the goal of achieving robustness, our research at UCR is directed towards learning parameters, feedback, contexts, features, concepts, and strategies of IU algorithms for model-based object recognition. Our multistrategy learning-based approach is to selectively apply machine learning techniques at multiple levels to achieve robust recognition performance. At each level, appropriate evaluation criteria are employed to monitor the performance and self-improvement of the system. We developed theoretically sound approaches to recognition and to learn segmentation for robust model-based recognition. We have developed two approaches based on reinforcement learning for closed-loop object recognition in a multi-level vision system. We show that in simple real scenes with varying environmental conditions and camera motion, effective low-level image analysis and feature extraction can be performed. We show the performance improvement of an IU system combined with learning over an IU system with no learning. Our initial research using outdoor video imagery and the Phoenix algorithm has demonstrated that (a) adaptive image segmentation can provide over 30 improvement in performance, as measured by the quality of segmentation, over non-adaptive techniques, and (b) learning from experience can be used to improve the performance over time. We have developed some novel techniques and we have some results for context reinforced ATR using learning techniques. These results have yet to be validated on a larger dataset.