Unsupervised saliency detection via kNN mechanism and object-biased prior

Abstract

Abstract In recent years, some researchers have defined the compactness hypothesis that similar colours tend to accumulate in the salient region rather than in the non-salient region for image saliency detection. Meanwhile, we found that the mechanism of k-nearest neighbour (kNN) also assumes that similar things exist in close proximity. Since the kNN method belongs to supervised learning, we only introduce the unfixed k-value and combine it with the clustering idea of k-mean to propose a new algorithm called kNN clustering. Based on a given compact matrix, an object-biased prior and an improved boundary and background prior are proposed. Our algorithm is extensively tested on three publicly available datasets, and the experimental results reveal that our algorithm obviously outperforms 19 existing saliency detection methods to agile generate the high-quality saliency maps with full resolution. At the same time, the improved prior methods are efficient with the existing algorithms without prior knowledge, especially for the low-performing models.