一种听觉显著图提取模型  

A saliency map extraciton model for auditory attention

作　　者：张欣[1] 夏秀渝[1] 王雪君[1] 

机构地区：[1]四川大学电子信息学院,成都610064

出　　处：《四川大学学报（自然科学版）》2014年第2期292-298,共7页Journal of Sichuan University(Natural Science Edition)

基　　金：国家自然科学基金(61071159)

摘　　要：本文基于人耳听觉特性提出了一种改进的自下而上听觉显著图提取模型.该模型根据人耳听觉系统对声音分频道处理以及声音信号时间方向的差异更能有效表征听觉显著性这两个特点进行显著图提取.首先对输入声音信号进行短时傅立叶变换(SFFT)得到声音信号的听觉图谱,然后重新划分频带并在各频带上采用一维高斯滤波器组进行不同尺度的时域滤波,再利用中心-周边差算子计算各频带的听觉显著度并进行跨尺度整和,最后将各频带的听觉显著度线性合并之后得到最终的听觉显著图.仿真实验的结果表明该模型提取的听觉显著图更准确.A new bottom-up saliency map for modeling auditory attention is proposed based on the char-acteristics of human auditory.Saliency maps can be effectively obtained by our model which is based on the two facts that auditory signal is frequency-divisionally processed by human auditory and spectrogram can be treated as a single dimensional temporal input at all times.First,the spectrogram is formed by applying the SFFT to the auditory signal.Then,the auditory spectrogram is divided into several fre-quent bands,and different sets of one-dimensional Gaussian filters are used to get different levels.The saliency maps at different scales at each band are achieved by center-surround mechanism.And by com-bining maps acorss scales,the saliency map for each specific frequency band is determined.The final sali-ency map for auditory attention is obtained by the linear combination of saliency map at each frequency band.The simulation results of auditory saliency map from the new model proposed are proven to per-form more accurately than the previous.

关 键 词：听觉显著图 显著图提取模型 听觉注意 

分 类 号：TN912.3[电子电信—通信与信息系统]