Abstract
Humans recognize objects effortlessly, in spite of changes in scale, position, and illumination. Emulating human recognition in machines remains a challenge. This paper describes computer vision algorithms aimed at helping visually-impaired people locate and recognize objects. Our neurally-inspired computer vision algorithm, called Attention Biased Speeded Up Robust Features (AB-SURF), harnesses features that characterize human visual attention to make the recognition task more tractable. An attention biasing algorithm selects the most task-driven salient regions in an image. Next, the SURF object recognition algorithm is applied on this narrowed subsection of the original image. Testing on images containing 5 different objects exhibits accuracies ranging from 80% to 100%. Furthermore, testing on images containing 10 objects yields accuracies between 63% and 96% for the 5 objects that occupy the largest area within the image subwindows chosen by attention biasing. A five-fold speed-up is attained using AB-SURF as compared to the time estimated for sliding window recognition on the same images.