Abstract
In the last decades, in order to make the processing of a scene more efficient, biologically inspired approaches have been proposed. Visual attention models are being studied and actively developed in order to reduce the complexity and computational time of the existing methods. We propose a biologically inspired model that combines a single pre-trained CNN architecture with an artificial foveal visual system that performs simultaneously the classification and localization of objects in images. This model is based on the fact that only a small part of the image is processed with high resolution at each time so we load a foveated image in the network and successively employ feed-forward passes to determine the class labels and then via backward propagation determine the object possible locations according to each semantic label. By directing the attention to the center of the proposed location we mimic the human saccadic eye movements. In the results obtained we used the ILSVRC 2012 validation data set in a GoogLeNet CNN. We demonstrate that for non-centered objects the gain of the classification performance between iterations is significant showing that when mimicking the human visual behaviour of foveation, saccades are needed to integrate the information at each time.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have