【摘要】 针对如何获得符合人类视觉感知的压缩图像问题，提出了基于条件生成对抗网络的图像压缩模型(HPIC).在HPIC中，首先利用一个超先验概率模型对原始图像进行编码量化，将条件附加标签和残差模块相结合的生成器用于压缩图像的重建，基于深度卷积神经网络搭建的判别器则用于区分压缩后的图像和真实图像间的差异.损失函数是基于比特率-失真-感知优化理论来设计的，一方面选用基于预训练Inception网络特征值的感知失真指标来实现具有高感知质量的图像压缩重建，另一方面利用生成对抗网络损失来消除压缩伪影，提高压缩精度.实验结果表明，HPIC在比特率-失真-感知三重权衡中取得了较好的平衡，即使目前的常见算法使用两倍于本文算法的比特率，本文算法在所有的感知指标得分上均优于前者，HPIC仍能够实现具有高感知质量的压缩.更多还原

【Abstract】 A conditional generative adversarial network-based high perceptual image compression(HPIC) model was proposed to obtain a compressed image that conforms to human visual perception. In HPIC, the original image was encoded and quantized based on a hyper prior probability model. A generator based on the conditional additional label and residual module was used for the reconstruction of the compressed image, and a discriminator based on the convolutional neural network was used to distinguish the difference between the compressed image and the real image. Based on the bit rate-distortion-perception optimization theory to design the loss function, the perceptual distortion metric was chosen based on the eigenvalues of the pre-trained inception network to achieve image compression reconstruction with high perceptual quality on the one hand, and use the adversarial generation network loss to eliminate compression artifacts and improve the compression accuracy on the other hand. The experimental results show that HPIC achieves a good balance in the bit rate-distortion-perception tradeoff and outperforms the current common algorithms in all perceptual metrics scores, even though the latter uses twice as much bit rate as ours. HPIC can achieve compression with high perceptual quality.更多还原