Image Saliency Detection In Complex Environments

With the rapid development of computer science, image processing is becoming more and more popular. Being an interdisciplinary of computer vision, psychology and neurobiology, image saliency detection has attracted much attention, and achieved great research progress in recent years. However, conventional methods usually obtain poor performance in complex environments. To handle this problem, this thesis investigates several solutions by establishing more robust saliency models and integrating multi-spectral information, including graph representation model with fusing low-level and high-level features and the edge-preserving filtering based multi-modal saliency model.The main works of this thesis are as follows:(1) A new graph representation model, which integrates low-level and high-level features, is proposed in this thesis for complex environment to cast the image saliency detection to the Markov random walk problem. Firstly, we take superpixels as graph nodes to construct the graph, in which the weights of the node and edge are defined by high-level features of node and the difference of low-level features between nodes respectively. Secondly, asymmetric transition probability matrix is constructed according to the proposed graph representation model and then, the Markov random walk algorithm is utilized to obtain the initial saliency map. Thirdly, the center prior and the improved boundary prior are employed to further improve the saliency map. Finally, extensive experiments on four publicly available datasets with ten approaches demonstrate the effectiveness of the proposed approach.(2) This thesis introduces thermal spectral information to overcome the limitations of visible spectral information in some complex environments, such as low illumination, haze, bad weathers. To this end, an edge-preserving filtering based multi-modal fusion model is proposed for efficient image saliency detection. Experiments on the newly created multi-modal dataset show that the proposed method can efficiently and effectively integrate multi-modal information to obtain robust saliency detection results.