基于无监督栈式降噪自编码网络的显著性检测算法

Saliency Detection Based on Unsupervised SDAE Network

针对现有的显著性检测算法检测目标类型单一、通用性差的问题,提出一种基于无监督栈式降噪自编码网络的显著性检测算法.该算法利用无监督栈式降噪自编码网络(Stacked Denoising Auto Encoder,SDAE)在多个尺度对原始图像进行稀疏重构,将原始图像与SDAE网络重构图像之间的差作为显著图,二值化后的显著图作为显著性目标检测结果.在SDAE网络训练过程中,将原始图像作为原始数据,网络重构的图像作为观察数据.为了提升网络训练效率,首先利用无监督逐层贪婪方法训练同结构的深度信念网络(Deep Belief Network,DBN),将训练得到的DBN网络参数设为SDAE网络的初始参数,再计算原始数据与观察数据之间的互信息作为网络收敛代价,利用反向传播进行网络参数微调.实验表明,该网络模型可以完成多类型目标的显著性检测,具有通用性好,准确度高等优点.

The traditional saliency detection method is difficult to detect different kinds of saliency target simultaneously.In order to solve this problem,an algorithm based on unsupervised SDAE network is proposed in this paper.The stacked denoising auto-encoder(SDAE) network is used to sparsely reconstruct original image in multiple scales.The difference between the original image and the reconstructed image is used as a saliency map,and the binaryzation of the saliency map is used as salient detection result.In the process of SDAE network training,the original image is used as the original data and the reconstructed images are treated as observed data.In order to improve the efficiency of network training,the deep belief network(DBN) is trained by greedy method in each layer without supervising,and the network parameters are delivered to stacked denoising auto-encoder(SDAE) network as initial parameters.Then,the mutual information between the original data and the observed data is used as loss function,and the network parameters are tuned by backpropagation.The experiments show that the proposed algorithm can accomplish the saliency detection of various targets,which has the advantages of good universality and high accuracy.