Visual saliency via loss coding

A novel and effective bottom-up saliency model inspired by the recent findings of the early vision system is proposed. The lossy coding length, which resembles the neural cost in the hierarchical structure of human vision system, is exploit to measure saliency. We show that the proposed efficient coding network can be considered as the coding process in the early vision system. The sparse coding process in simple cells of the primary visual cortex and a dimensionality reduction process via the principal component analysis are integrated in the proposed network. The saliency value at each image pixel is computed based on the residual of the coding process. The proposed biological-inspired saliency model is evaluated on two different eye-tracking datasets against several state-of-the-art algorithms. Experimental results demonstrate the effectiveness, efficiency as well as robustness of the proposed model, and bear out the hypothesis of lossy coding for visual saliency.