Nonnegative Matrix Factorization-based Audio Restoration Method For Historial Film

Historical film materials suffered from audio noise caused by deterioration and breakage of films during a long term storage,which leads to low quality and poor auditory sensation of audio sound.Thus it's necessary to restore historical film since it belongs to historical relic.Currently,there are numbers of restore methods in relative field,but as for background noise,there are still no feasible restoration approaches in film industry.A Non-negative matrix factorization(NMF)-based audio noise restoration method is proposed to solve the above problem in this paper.In the proposed method,firstly,a number of noise samples which taken from historical film are trained to obtain the priori noise information,then the NMF algorithm can separate the noise and non-noise component by using trained priori noise feature,thus the noise restoration is implemented.The paper mainly contains following works:NMF-based historical film audio noise restoration workflow is proposed,which divided into training phase and restoration phase.In training phase,noise samples in historical film are trained by using NMF to build the noise model.In restoration phase,considering trained noise model as input in NMF,the separation of noise is implemented by NMF.In order to solve the problem of non-noise signal distortion after NMF-based restoration,a linear constrained and sparse representation modified NMF algorithm is proposed to replace the original NMF algorithm used in restoration phase for noise separation.Linear constraints are carried out for non-noise feature by sinusoidal model,and sparse representation for whole signal feature are made in modified NMF.The experimental results show that modified NMF can effectively restore the noise,meanwhile it also decreases the distortion level of non-noise signal compared to original NMF algorithm.The proposed modified NMF algorithm is applied in historical film audio noise restoration,and it has been compared with iZotope and Sonnox denoise perform algorithms by using Signal to Noise Ratio(SNR)as quality objective measure.The experimental results show that the proposed modified NMF algorithm outperforms others.