基于图像语义的弱监督显著性物体检测

Weakly Supervised Salient Object Detection Based on Image Semantics

为了减少显著性物体检测对像素级标签的依赖,提出了一种基于图像语义的弱监督显著性物体检测方法.利用鱼网络和注意力机制的组合模型,在图像语义热力映射图的基础上,对弱标签采用余弦相似度进行训练更新,同时在网络训练初期采用训练诱导策略,利用简单数据集对整个网络进行诱导训练,使其具有一定的能力.然后,经过不断地增加数据集的复杂性,使得网络提取特征的能力越来越强.在4个显著性检测数据集上进行实验,并与传统监督方法进行对比分析,实验结果表明,该方法的F-MAX值在各个数据集上平均提高0.03~0.08,MAE减少0.02~0.05,在较弱的监督标签下能更精确地提取图像中的显著性特征.

To reduce the dependence of salient object detection on pixel-level labels,we propose a weakly supervised salient object detection method based on image semantics.Using the combined model of the fish network and the attention mechanism,on the basis of the image semantic heat map,the weak labels were trained and updated by cosine similarity.At the same time,we used a training induction strategy.In the initial stage of network training,a simple dataset was used to induce the entire network to make it have certain capabilities.Then,after continuously increasing the complexity of the dataset,the network’s ability to extract features became stronger and stronger.Experiments were conducted on four saliency detection data sets,and compared with traditional supervision methods.The experimental results show that the F-MAX value of this method is increased by 0.03‒0.08 on each data set on average,and the MAE is reduced by 0.02‒0.05.Under the weak supervision label,this method can more accurately extract the salient features.