基于全局和局部信息融合的显著性检测

Saliency detection via fused global and local information

为提高低对比度、复杂自然图像显著性检测的准确率和泛化性能,提出一种贝叶斯框架下的全局和局部信息融合的显著性检测模型.首先,构建深度卷积自编码网络,采用对称编解码结构,监督学习图像全局特征,得到全局显著图;然后,根据全局显著图产生前景和背景码本,利用局部约束线性编码算法进行编码,采用稀疏编码描述局部特征,产生局部显著图;最后,提出采用贝叶斯框架,将全局和局部信息融合,生成最终显著图.实验结果表明,所提模型在ECSSD,DUT-OMRON和PASCAL数据集上F-measure值分别为76.53%、59.45%和72.52%,MAE值分别为0.14328、0.13787和0.18105,且能够有效对低对比度、复杂真实自然图像进行显著性检测.

Aiming at improving the accuracy and generalization ability of natural images with low contrast and complex background,a saliency detection model based on Bayesian framework is proposed via fusing global and local information.Firstly,a deep convolution autoencoder network was constructed,and symmetrical encoder and decoder structure was adopted to supervise learning the global features,and global saliency map was generated;Secondly,foreground and background codebooks were obtained from the global saliency map,and the locality-constrained linear coding method was used for encoding,and the sparse coding was employed to describe the local features,so local saliency map was generated;Finally,the Bayesian framework was adopted to integrate the global with local information,and the final saliency map was thus obtained.The experimental results show that the F-measure of our model on ECSSD,DUT-OMRON and PASCAL datasets is 76.53%,59.45%and 72.52%respectively,and the MAE is 0.14328,0.13787 and 0.18105 respectively.Furthermore,our method can effectively detect the salient object in real natural images with the low contrast and complex background.