Abstract
Saliency detection becomes more and more popular in computer vision research field. In this thesis we present a new method to generate the saliency map. 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. Our method consists of two parts: training step and testing step. In the training step, we use the features generated from images and the fixation values from ground-truth fixation map to train the feature-based dictionary for the sparse coding and the fixation-based dictionary for converting the sparse coding to a saliency map. In the test step, given a new image, we can get its corresponding sparse coding from the feature-based dictionary and then generate the result. We evaluate our results on two datasets with the shued AUC score and demonstrate that our method gives an efficient sparse coding learning and combination for saliency detection.