Cognitive Mechanisms Underlying the Classification of Reduced Dimensionality, Information Rich Image Representations

Both humans and computers exhibit significantly higher accuracy in natural scene classification when images are represented in a reduced dimensionality, information rich (RDIR) format compared to a blurred and downsampled (DS) format. RDIR representations are generated by processing the original image with an algorithm that captures prominent orientation information in the scene, inspired by the way humans capture the gist of a scene upon observing it for 200 milliseconds or less. DS scenes are generated by blurring the image with a Gaussian filter and then resizing the image to the desired resolution (a method currently used in biomedical devices that partially restore scene understanding for visually impaired users). A possible cognitive mechanism, inspired by human color vision, is proposed for the enhanced recognition accuracy observed with RDIR representations based on the higher decorrelation of image pairs depicted in RDIR format compared to that of their DS counterparts, drawing an analogy to the decorrelation observed in the first two principal components of human red and green cone responses.