Saccadic Object Recognition using a Markov Decision Process in a Cascaded Framework

Important issues in sequential object recognition are to define a strategy for saccadic access of visual information, and the representation of the features under observation. Selective attention has been described from the viewpoint of decision behaviour in experimental psychology (eg Gorea and Sagi, 2003 Perception 32 Supplement, 5) and in computer vision within the frame-work of Markov decision processes (MDPs) [Paletta and Pinz, 2000 Robotics and Autonomous Systems 31 71-86; Minut and Mahadevan, 2001 Proceedings of the Fifth International Conference on Autonomous Agents (Montréal: ACM Press) pp 457-464]. The original contribution of this work is to embed the saccadic decision procedure in a cascaded recognition process where visual evidence is probed exclusively at salient image locations. In a first processing stage, salient points are determined from an entropy-based cost function on object discrimination. Local information in terms of code-book vector responses determines the recognition state in the MDP. A reinforcement learner performs then trial-and-error search on useful actions towards salient locations within a neighbourhood, receiving reward from entropy decreases. After training, the decision maker demonstrates feature grouping by hypothesis verification behaviour. The method is evaluated in experiments on object recognition with the COIL-20 database, proving the method being computationally feasible and providing high recognition rates.