弱监督学习下的视觉显著性目标检测算法

摘要

为模拟人类视觉对含有特定目标图像集中目标逐渐关注感知的行为,提出一种弱监督学习的视觉显著性目标检测算法.根据已有的视觉显著性方法获得图像的显著性区域;提取显著区域的底层视觉特征,训练获得视觉显著目标的表征;用条件随机场(conditional random fields,CRF)将学习到视觉显著目标表征进行联合学习,获得该表征在最后显著性中的权重;计算每次迭代显著图的ROC曲线,寻找视觉显著性目标最优表征及其在最后显著图中的最优权重.实验结果表明,该算法检测精度优于现有诸多算法,能够有效检测出视觉显著性目标.该算法模拟了人类视觉中对特定关注目标的感知过程,对不断重复出现的视觉显著性目标进行强化学习,具有较高的准确率.%Aiming at simulating a human visual sense that people will gradually focus on specific object in a target image set, a visual salient object detection via weakly supervised learning was proposed.According to the state-of-art saliency method, the saliency regions of image were obtained.The low-level visual feature of saliency regions was extracted, and it was used to train appearances of visual salient object.A conditional random fields (CRF) model was built to learn the model coefficient together with the appearances of saliency object.The area of ROC was calculated after each iteration so as to obtain the best appearances of vi-sual saliency object and the weight of it in the final saliency map.Experimental results on the dataset indicate that this method performs much better than the existing state-of-art approaches, and it can detect the visual saliency object efficiently.This me-thod simulates human visual sense procession, the repetitive visual saliency object can be learnt and emphasized, and at the same time it has good accuracy.