Saliency Detection Using Quaternion Sparse Reconstruction

We proposed a visual saliency detection model for color images based on the reconstruction residual of quaternion sparse model in this paper. This algorithm measures saliency of color image region by the reconstruction residual and performs more consistent with visual perception than current sparse models. In current sparse models, they treat the color images as multiple independent channel images and take color image pixel as a scalar entity. Consequently, the important information about interrelationship between color channels is lost during sparse representation. In contrast, the quaternion sparse model treats the color image pixels as a quaternion matrix, completely preserving the inherent color structures during the sparse coding. Therefore, the salient regions can be reliably extracted according to quaternion sparse reconstruction residual since these regions cannot be well approximated using its neighbouring blocks as dictionaries. The proposed saliency detection method achieves better performance on Bruce-Tsotsos dataset and OSIE dataset as compared with traditional sparse reconstruction based models and other state-of-art saliency models. Specifically, our model can achieve higher consistency with human perception without training step and gains higher AUC scores than traditional sparse reconstruction based models.