采用渐进式网络的弱监督显著性目标检测算法

Weakly-supervised salient object detection with the multi-scale progressive network

弱监督显著性目标检测中常存在目标错检、区域检测不全和目标边界不清晰等问题。针对上述问题,提出了一种基于渐进式网络的弱监督显著性目标检测算法,将显著性目标检测分为目标定位、显著性区域完善和目标边界细化3个子任务分阶段完成。首先,将输入图像采样为3个不同尺度的图像,分别输入渐进式网络的3个阶段进行学习;其次,在目标定位阶段设计了嵌套位移多层感知机,平衡网络的全局信息与局部信息的提取能力,以更好地定位显著性目标;最后,根据显著性图的结构不受尺度变化影响的特点,设计了异尺度自监督模块和目标一致性损失函数来构建自监督机制,使网络能够输出区域完整、边界清晰的显著性图。在5个数据集上测试所提算法,其客观指标与主观评价都优于最近的弱监督算法,且在F值指标上可以达到相关全监督算法89%的性能。实验结果表明,所提算法能生成显著性区域更完整、显著性目标边界更锐利的显著性图,且具有良好的鲁棒性。

Existing weakly-supervised salient object detection methods often suffer from problems such as false positive,low recall rate,and unclear edges.To address the above issues,a weakly-supervised salient object detection with the multi-scale progressive network is proposed,which divides salient object detection into three sub-tasks:object localization,saliency region improvement and edge refinement.First,the input image is sampled into three images of different scales,which are respectively fed into the three stages of the multi-scale progressive network for learning.Second,in order to better locate the salient objects,a nested shift multi-layer perceptron is proposed in the object localization stage,which can balance the global feature and local feature extraction ability of the network.Finally,according to the characteristic that the structure of saliency maps is not affected by scale changes,a multi-scale self-supervision module and an object consistency loss are designed to build a self-supervision mechanism,so that the network can output a saliency map with complete regions and sharp edges.The proposed method is tested on five datasets,and outperforms the recent weakly-supervised methods in both quantitative and qualitative comparisons,and can reach 89%of the performance of the related fully-supervised methods on the F-measure index.Experimental results show that the proposed algorithm can generate saliency maps with complete saliency regions and sharp edges,and has good robustness.