基于改进生成式对抗网络的水下图像增强

Underwater Image Enhancement Based on Improved Generative Adversarial Networks

摘要:

由于水介质的吸收和散射特性会导致雾化、低对比度、颜色退化等各种水下成像失真，严重影响了水下图像的后续利用。为恢复清晰的水下图像，提出一种基于改进生成对抗网络的深度学习模型。借助图像质量评价技术，将生成的过程样本与高质量样本进行拟合，并将拟合得到的差值信息用于优化网络中的生成器。改进的生成式对抗网络有效改善了由真假训练逻辑带来的图像质量提升限制的问题。实验结果显示，该方法有效的恢复了水下图像的色彩，并改善了图像的清晰度和对比度；相比其他方法，提出的方法在SSIM、UCIQE和UIQM指标上分别提升了2.9%、6.2%和14.3%。

Abstract:

Due to absorption and scattering in the underwater environment, underwater images often suffer from fogging, low contrast and color degradation, which seriously affect the subsequent tasks. To obtain clear underwater images, an improved GAN-based deep learning method is proposed for underwater image enhancement. By employing the image quality assessment technique, the fitting information of generated intermediated samples and high-quality samples is applied to optimize the generator of the proposed network, which improves the limitation of the true-or-false training strategy. The experimental results show that the proposed method can effectively improve the clarity and contrast, and restore the color of the underwater images. In addition, comparing with several state-of-the-art methods, our proposed methods achieve increase of 2.9%, 6.2% and 14.3% in terms of SSIM, UCIQE and UIQM indexes, respectively.