Saliency detection using tensor sparse reconstruction residual analysis

In this paper, a visual saliency detection model based on tensor sparse reconstruction for images is proposed. This algorithm measures saliency value of image regions by the reconstruction residual and performs better on color images than current sparse models. Current sparse models treat a color image as multiple independent channel images and vectorise the image patches, ignoring interrelationship between color channels and spatial correlation between neighbouring pixels. In contrast, the proposed tensor sparse model treats a color image as a 3D array, retaining the spatial color structures entirely during the sparse coding. The proposed saliency detection method is tested on ASD dataset and OSIE dataset and compared with traditional sparse reconstruction based models. The experimental results show that our model achieves higher AUC scores than traditional sparse reconstruction based models.