Recovery of Lossy Compressed Music Based on CNN Super-Resolution and GAN
Yandong Liu · 2021 IEEE 3rd International Conference on Frontiers Technology of Information and Computer (ICFTIC) · 2021
The lossy compression of music audio files through MP3 cause loss of sound quality, which results in the decline of auditory experience and cannot meet the requirements of high-quality music playback in a wide variety of occasions. To solve this problem, after long-term exploration, we proposed an approach of time-domain and frequency-domain bandwidth expansion based on CNN and GAN to support high-quality MP3 lossy compressed music recovery by analyzing the characteristics and correlation of human voice and different kinds of musical instruments in high-frequency and low-frequency portions of music. For bandwidth extension in the frequency domain, a method similar to image inpainting is designed; for bandwidth extension in the time domain, a super-resolution method is designed. We compared the proposed method with RNN, BPCNN, and other methods. The experimental results prove that the method proposed in this paper has the lowest spectral loss and the best reconstruction quality. The experimental results of human ear discrimination further prove the effectiveness of the audio enhancement algorithm.