A salient object detection algorithm based on convex hull calculation and wavelet transform

DOI:	10.11809/bqzbgcxb2023.04.036
Keywords:	saliency detection; superpixel segmentation; convex hull calculation; wavelet transform; multi scale
Abstract:	Aiming at the problems of the traditional salient detection algorithm, such as inaccurate detection of saliency regions, insufficient brightness of regions and ineffective suppression of background interference, this paper herein proposes a salient object detection algorithm based on convex hull and wavelet transform (CHWT). Firstly, the convex hull intersection is obtained by calculating the convex hull of the input images in RGB, Lab, and HSV space respectively, and then the binarization mask is performed on the convex hull intersection. Secondly, the input images are superpixel segmented at multiple scales, and the saliency graphs at multiple scales are obtained by using manifold ranking (MR) algorithm. Through the combination with the saliency graphs at multiple scales and Bayesian fusion with the binarization mask of the convex hull intersection, the multi scale superpixel convex hull saliency graphs are obtained. Finally, wavelet transform is used to decompose the DCT coefficient amplitude spectrum of the input images from multiple dimensions, and the multi scale wavelet transform saliency graphs are obtained. Besides, final saliency graphs are achieved through the linear fusion with the multi scale superpixel convex hull saliency graphs. The results show that, compared with the performance of the six other existing algorithms on MSRA 10k, ECSSD and HKU IS datasets, the proposed CHWT algorithm in this work is superior in indicators like the Precision Recall curve (PR curve), algorithm comprehensive index (F measure), mean absolute error (MAE) and structure measurement (SM) indexes, which shows better stability and robustness of the proposed CHWT algorithm.
Issue:	Vol. 44 No. 4 (2023)
Published:	2023-04-28
PDF HTML