Learning sparse dictionaries for saliency detection

We present a new method of predicting the visually salient locations in an image. The basic idea is to use the sparse coding coefficients as features and find a way to reconstruct the sparse features into a saliency map. In the training phase, we use the images and the corresponding fixation values to train a feature-based dictionary for sparse coding as well as a fixation-based dictionary for converting the sparse coefficients into a saliency map. In the test phase, given a new image, we can get its sparse coding from the feature-based dictionary and then estimate the saliency map using the fixation-based dictionary. We evaluate our results on two datasets with the shuffled AUC score and show that our method is effective in deriving the saliency map from sparse coding information.