A Bottom-up Saliency Detection Method
Xiaosong Zhou, Ping Zeng · 2021
A bottom-up saliency detection method is presented to handle image in this paper. An image is first filtered by a bandpass filter to decide the seeding region, and k-means with initial seeds selected from the seeding region is applied to extract several low-level image features. Next seven image statistics are computed using pixels of each cluster, and entropy is calculated for each feature in each cluster. Then a unique feature (so-called entropy order sum) is defined by adding the contribution of each individual feature, and the relation between entropy order sum and saliency is investigated. Finally, salient clusters are detected by defining a hit rate of saliency between the entropy order sums of the various clusters.