基于联合生成-去除的水下图像增强方法

Joint Underwater Image Generation and Removal via Disentangled Image Translation

摘要:

现有的基于深度学习的水下图像增强方法在仿真的水下图像上取得了良好的效果。但是,由于简化的仿真图像与复杂的真实图像之间存在较大差距,此类方法在处理真实水下图像时性能明显下降。为了解决真实水下图像增强问题,提出了一种联合生成-去除水下图像增强方法。该方法采用分解思路,将水下图像分解为干净的背景层和退化层,通过循环一致性损失和对抗性损失来更好地保留背景,进而实现真实图像和仿真图像之间的转换,既校正了图像颜色,又提升了图像对比度,实现良好的增强效果。实验结果表明,本方法在真实水下图像数据集上处理的结果,在色彩、纹理细节和清晰程度方面均优于现有的对比方法。

Abstract:

The existing underwater image enhancement methods based on deep learning have achieved good results in synthetic underwater images. However, due to the large gap between the simplified synthetic image and the complicated real image, the performance of learning based methods will be significantly degraded when dealing with real images. To solve the problem, a joint underwater image generation and removal (JUIGR) method is proposed. The underwater image is decomposed into a clean background layer and a degraded layer, and the background is restored through cycle consistency loss and adversarial loss. Moreover, the transformation between the real and synthetic image further corrects the color and improves image contrast to achieve a good enhancement effect. Extensive experimental results show that the proposed method is superior to the existing methods in terms of color, texture detail and clarity on the real underwater image dataset.