Traditional algorithms for solving single-channel blind source separation tasks often rely on prior information about multiple independent source signals,such as low-rank,sparsity,and temporal continuity.In this paper,we propose a novel SCBSS algorithm based on GAN.This algorithm aims to overcome the reliance on prior information.Utilizing the generator of GAN,the algorithm roughly separates the source signals and innovatively integrates distribution constraints,energy balance,and adversarial constraints to ensure effective separation without any prior information.Experimental validation on publicly available handwritten digit dataset MNIST images demonstrates the good performance of the proposed algorithm.Performance evaluation using two key metrics of SCBSS,peak signal-to-noise ratio and structural similarity,confirms the effectiveness of image separation.Results show that the proposed algorithm outperforms some commonly used prior-based traditional algorithms and leverages its unsupervised advantage for applications in restoring ancient Chinese characters.
关键词
SCBSS/GAN/对抗约束/汉字修复
Key words
SCBSS/GAN/adversarial constraints/Chinese character restoration
万方数据