基于视觉质量驱动的无监督水下图像复原

Unsupervised Underwater Image RestorationBased on Visual-Quality-Driven

基于深度学习的水下图像复原方法已经取得了较大进展,但大多数是有监督学习方法,其通常需要大量高质量的清晰图像作为参考,难以应用于实际水下场景.因此,从水下图像降质机理和视觉质量提升等角度出发,提出一种基于视觉质量驱动且无需任何参考图像的无监督水下图像复原方法.网络主体采用基于小波变换的编解码结构,包含一些对称的卷积层和反卷积层,学习端到端的数据映射;为了实现特征复用,在编解码对应层加入跳层连接,增强特征提取能力;使用小波变换取代传统上采样和下采样,保证特征提取过程中较好地保持图像细节;设计残差增强模块,有效缓解梯度消失和过拟合现象.更为重要的是,从图像视觉质量改善角度出发,设计一系列无参考损失函数,包括感知损失、全变差正则化损失、颜色损失、对比度损失等,网络可实现由水下降质图像到清晰图像的直接输出.实验结果表明,所提无监督水下图像复原方法优于部分监督式方法,无参考评价指标Entropy、UIQM分别取得最高值7.38、3.31,且复原图像细节丰富、色彩自然,具有较强的泛化能力,可较好用于实际水下场景.

Deep learning has achieved quite promising results in underwater image restoration.However,most of them are supervised learning methods,which typically require massive high-quality images as reference images and cannot be applied to practical underwater settings.Therefore,we proposed a visual-quality-driven unsupervised underwater image restoration method from the perspectives of underwater image degradation mechanism and visual quality improvement.The backbone network adopts the encoder-decoder structure based on a wavelet transform and learns an end-to-end data map,which includes some symmetric convolutional and deconvolutional layers.To enable feature reuse,a skip connection was added to the corresponding layers of the codec to enhance the feature extraction capability.We used wavelet transform to replace traditional up-sampling and down-sampling to preserve more details in the feature extraction process.In addition,the residual enhancement module was designed to effectively relieve gradient disappearance and overfitting phenomena.More importantly,we designed a series of no-reference loss functions,including perceptual loss,total variation regularization loss,color loss,and contrast loss,to restore the visual quality of underwater images.The network can realize the direct output from the degraded images to sharp ones.The experimental results show that the proposed unsupervised underwater image restoration method outperforms partially supervised methods.The no-reference evaluation indices Entropy and UIQM achieve the highest values of 7.38 and 3.31,respectively,and the restored image has rich details and natural colors.It has a strong generalization capability and can be better used in real underwater scenarios.