Adaptive Spatiotemporal Feature Extraction and Dynamic Combining Methods for Selective Visual Attention System

摘要

This paper introduces a selective attention system that guides users to detect the regions of interest more effectively by adaptively selecting and using spatial and temporal features according to the input images. Although the proposed system is based on a typical bottom-up method, it achieved improvement in the method for extracting features and calculating the saliencies compared to existing studies. In the proposed system, spatial saliencies have dynamic information from which features are adaptively selected according to the input images. Also temporal saliencies in the proposed system have pieces of information for individual moving objects that are associated with each other obtained through multi-resolution feature analysis. In addition, when combining a spatial saliency and a temporal saliency, the activity of the input saliency is measured, and the weights that change dynamically according to the activity are calculated, and the spatial saliency and temporal saliency are combined according to the weights. In order to evaluate the performance of the proposed system, comparative experiments with the existing systems were conducted with diverse experimental images and as a result, it could be seen that the proposed system produces results closer to the results of humans' visual recognition compared to previous systems.