Cascaded Sequential Attention for Object Recognition with Informative Local Descriptors and Q-learning of Grouping Strategies

The contribution of this work is to provide a three-stage architecture for sequential attention to provide a system being capable of sensorimotor object detection in real world environments. The first processing stage provides selected foci of interest in the image based on the extraction of information theoretic saliency of local image descriptors (i-SIFT). The second stage investigates the information in the local attention window using a codebook matcher, providing local weak hypotheses about the identity of the object under investigation. The third stage then proposes a shift of attention to a next attention window. The working hypothesis is to expect a better discrimination from the integration of both the individual local FOA patterns and the geometric relation between them, providing a model of more global information representation, and feeding into a recognition state in the Markov Decision Process (MDP). A reinforcement learner (Q-learner) performs then explorative search on useful actions, i.e., shifts of attention, towards locations of salient information, developing a strategy of useful action sequences being directed in state space towards the optimization of discrimination by information maximization. The method is evaluated in experiments using the COIL-20 database (indoor imagery) and the TSG-20 database (outdoor imagery) to demonstrate efficient performance in object detection tasks, proving the method being more accurate and computationally much less expensive than standard SIFT based recognition.