亮度自调节的无监督图像去雾与低光图像增强算法研究

Study on Unsupervised Image Dehazing and Low-light Image Enhancement Algorithms Based on Luminance Adjustment

在低质图像降质问题中,亮度偏离(如图像偏亮及偏暗)是较为常见的图像降质现象。基于全监督学习的图像增强方法面临训练数据难以获取或获取成本过高、训练数据和应用场景不一致的困境。针对以上问题,提出了一种能够克服数据依赖和亮度自适应的无监督图像增强方法。方法的具体细节为:针对图像去雾与低光增强任务,设计了一个基于通道与像素注意力的深度卷积网络,对增强图像与输入图像进行比较,采用亮度饱和度、空间一致性、照明平滑度、伪标签监督损失等多种无监督损失函数,在保证增强图像与输入图像一致性的同时,调节图像的亮度偏离程度,有效提高图像质量。实验结果表明,所提方法在客观指标及视觉效果上不仅优于传统方法和基于无监督学习的方法,甚至优于近年来的全监督图像增强方法。将所提方法与5种图像去雾方法及4种低光增强方法分别进行对比,相比性能次优的方法,其在图像去雾任务的Reside数据集上,PSNR和SSIM值分别提高了2.8 dB和0.01;在低光增强任务的SICE数据集上,PSNR和SSIM分别提高了0.56 dB和0.01。结果表明,所提无监督图像去雾与低光增强算法能够有效调节图像的亮度偏离程度,重建了亮度正常、细节清晰、对比度较好的增强复原图像,较为有效地克服了当前底层视觉任务数据难以获取、训练数据与应用数据不一致、存在域迁移的难题,提升了算法在应用中的适用性。

Among the degradations of low-quality images,luminance deviations such as brighter or darker images are very common image degradation phenomena.The image enhancement method based on fully supervised learning faces the dilemma that the training data is difficult to obtain or the acquisition cost is too high,and the training data is inconsistent with the application scene.To handle these problems,an unsupervised image dehazing and low-light enhancement algorithm based on luminance adjustment is proposed in this paper.A deep architecture with channel attention and pixel attention mechanism are designed to measure the differences between enhanced images and input low-quality images.A variable quadratic function has been applied to adjust the pixel luminance of the image.Multiple unsupervised losses i.e.,brightness saturation loss,spatial consistency loss,illumination smoothness loss and pseudo-label supervision loss are utilized to alleviate the illumination deviations but to ensure the identity between the enhanced images and the input low-quality images,which efficiently improve the quality of the images.Empirically,an intensity compression strategy is applied for the hazy images to darken the hazy images to have a similar intensity range with low-light images.Thus,the hazy images can be treated equally with low-light images with our deep network to adjust the luminance of the image.For the dehazing task,compared with the second-best method,our method improves the PSNR value for 2.8 dB and SSIM value for 0.01 in RESIDE dehazing dataset.For the low-light enhancement task,our method outperforms the second-best method for 0.56 dB and 0.01 separately measured by PSNR and SSIM in the SICE dataset.The proposed image dehazing and low-light enhancement algorithms can restore high-quality images from hazy images and low-light images.It effectively overcomes the difficulty of acquiring the targeted enhanced data or alleviates the problem of domain gap between training data and application data in the low-level vision tasks,which improves its adaptivity in real applications.