关键词: 关键词: 显著性目标检测, Frequency-tuned算法, Grab cut算法, 可微分聚类, 由“粗”到“精”

Abstract: Abstract: For the low segmentation accuracy of traditional saliency detection algorithms and the strong dependence on pixel-level manual annotation data of deep learning algorithm, an unsupervised saliency target detection method based on graph cut refinement and differentiable clustering was proposed. The idea of "coarse" to "fine" was used for realizing accurate saliency target detection which only depended on the characteristics of a single image. Firstly, the frequency tuned algorithm was used for getting the salient coarse image according to the color and brightness contrast information of the image itself. Then, the candidate regions of the salient target were obtained by binarization according to its statistical characteristics and the central priority hypothesis. Then, the Grab cut algorithm based on single image graph cutting was used for segmenting the salient target. Finally, in order to overcome the difficulty of imprecise detection when the background was very similar to the target, the unsupervised differentiable clustering algorithm with good boundary segmentation effect was introduced to further optimize the saliency map. The experimental results show that compared with the existing seven algorithms, the optimized saliency map of the proposed algorithm is closer to the ground truth, and the Mean Absolute Error of 14.3% and 23.4% can achieve on ECSSD and SOD datasets respectively.

Key words: salient object detection, Frequency-tuned algorithm, Grab cut algorithm, differentiable clustering, from coarse to fine