Saliency Detection Based On Multi-scale Superpixel And Dictionary Learning

In recent years, digital image processing technology is more and more widely applied to various fields, including transportation, security, medical, etc. In intelligent transportation, biological certification and computer aided medical, it has a broad development space and market application prospect. But, how the computer can be fast like a human to understand the meaning and content of image has now become a subject to the majority of researchers. Digital image contains huge amounts of data, the big data processing has become the hot issues in recent years. The technology for saliency detection can reduce redundant information and highlight important areas, which has an important advantage in terms of reducing the capacity of big data. Therefore, the research of visual saliency detection has become an important content in the field of digital image processing.Physiological research results show that the pathways in the brain, which transfer visual information, present a complex hierarchy structure. While biological visual cortex has the learning mechanism, the receptive field of single cell produces sparse representation. So, based on the hierarchical visual computing theory, we propose a visual saliency detection algorithm based on multi-scale superpixel to simulate the human visual perception mechanism in this thesis. Because size increasing gradually is the basis of hierarchical visual calculation, we detect the saliency of image in more than one superpixel scale. Human physiology structure accords with sparse representation and learning mechanism. Thus, the method learns a background dictionary based on sparse coding and dictionary learning, and calculates the sparse coefficient of each superpixel. Salient objects are different from the background and unique. So, according to reconstruction error of each pixel, the saliency map is generated corresponding superpixel scale. And then, the different superpixel scale saliency maps fuse together to generate the final saliency map. The saliency map is hierarchical, which can make up the weakness that the performance in single layer is unstable.In order to verify the performance of this algorithm for saliency detection, called based on multi-scale superpixel and dictionary learning, we carry on the contrast experiment with other five algorithms, which are representative and have good detection performance, on four databases, including CSSD, ECSSD, DUT-OMRON and CMU. The experimental results show that the proposed algorithm can highlight the object region inhibiting background and be better at precision, recall and F-measure with comparative ones.