基于注意力机制的深度学习图像压缩

Deep Learning Image Compression Based on Attention Mechanism

图像压缩是一个基础性的研究领域,许多压缩标准已经发展了几十年。最近,基于卷积神经网络的图像有损压缩逐渐取得一系列显著的进展。目前,最有效的基于学习的图像编解码器采用自动编码器的形式,采用了通道调节(CC)和潜在残差预测(LRP)来提高压缩性能,但图像仍然存在空间上的冗余,从而影响到率失真性能。为了使这一问题得到改善,文章提出使用RBAM注意力模块融入网络体系结构中,以提高性能。实验结果表明,用峰值信噪比(PSNR)作为评价指标,文章所提出的网络结构优于传统方法,达到了更好的率失真性能。

Image compression is a fundamental research field,and many compression standards have been developed for decades.Recently,image lossy compression based on Convolutional Neural Networks has made a series of remarkable progress.At present,the most effective image codec based on learning uses the form of automatic encoders,which uses Channel Conditioning(CC)and Latent Residual Prediction(LRP)to improve compression performance,but the image still has spatial redundancy and the rate-distortion performance is affected.To remedy this problem,this paper proposes to integrate the RBAM attention module into the network architecture to improve performance.The experimental results show that the proposed network structure is superior to the traditional method by using the Peak Signal-to-Noise Ratio(PSNR)as the evaluation index,and it achieves better rate-distortion performance.