Saliency Driven Nonlinear Diffusion Filtering for Object Recognition

We propose the saliency driven nonlinear diffusion filtering as a boost for object recognition. Taking saliency image as mask for magnitudes of gradients, nonlinear diffusion filtering treats foreground and background selectively. It preserves foreground information while filters out background information as much as possible. In salient area, semantically important structures are well preserved, while in non-salient area, cluttered structures are inhibited and smoothed into plain regions. Object recognition is conducted utilizing Bag-of-Words model, which can implicitly emphasize important foreground features for the reason of selective filtering. Experiments show that recognition accuracies using filtered images are generally higher than those using initial images, and are comparable with state-of-the-art. Consequently, we draw a safe conclusion that saliency driven nonlinear diffusion filtering undoubtedly help improve recognition performance, as long as saliency images are appropriate.