基于特征解耦表征学习的无监督混合失真图像复原方法

Unsupervised Hybrid-Distorted Image Restoration Method Based on Feature Disentangled Representation Learning

针对真实场景下多种混合失真组合的多任务图像复原,结合生成对抗网络与编码器提出了一种基于无监督对偶学习的图像复原方法.该方法引入特征解耦模块,通过修正基于增益控制的归一化,将不同退化机制的特征表示分配到不同的特征通道中,使得不同退化特征表达相互独立,实现了通道上的特征解耦.同时,为了进一步过滤掉退化机制的特征表示并保持原图像内容信息的细节,利用通道注意力机制实现特征解耦后自适应地选择有用的特征表示,使其适用于真实场景下混合失真组合的图像复原任务.实验结果表明,所提算法在单一退化类型GoPro数据集上对比基于尺度循环网络的算法(scale-recurrent network,SRN),峰值信噪比和结构相似性2项指标值分别提高了0.499 dB和0.044;在混合退化类型DIV2K数据集上对比基于操作选择注意力网络的算法(operation-wise attention network,OWAN),峰值信噪比和结构相似性2项指标值分别提高了0.163 dB和0.015,在实现复原图像的同时,保留了图像的纹理和细节信息.

For the hybrid-distorted image restoration in real scenes,an image restoration algorithm based on unsupervised dual learning is proposed with the generative adversarial networks and the encoder.The algo-rithm introduces feature decoupling module where the different feature representations from different deg-radation mechanisms are assigned to different feature channels by modifying the normalization based on gain control so the feature disentanglement in channels is realized by independent feature representations.Meanwhile,the channel attention mechanism is used to realize the restoration of image content feature of clean images.Compared with the scale-recurrent network(SRN)algorithm,the proposed algorithm im-proves 0.499 dB and 0.044 in terms of peak signal-to-noise ratio(PSNR)and structural similarity(SSIM)on GoPro dataset.Compared with the OWAN algorithm,the proposed algorithm improves 0.163 dB and 0.015 in terms of PSNR and SSIM on DIV2K dataset.The experiments also demonstrate that the detailed infor-mation can be restored by the proposed algorithm.