基于BcGAN的水下图像增强方法

Underwater image enhancement method based on BcGAN

针对目前水下图像增强的结果中存在雾度残留以及细节模糊的问题,提出一种基于条件生成对抗网络(cGAN)的Boosting水下图像增强方法 (BcGAN)。在编码器-解码器结构的基础上引入SOSboosting策略得到增强生成器,实现对图像特征的逐步细化;提出双重判别器结构,实现对多尺度输入的判别,以WGAN-GP损失为主导构建双重判别器的联合训练损失函数。实验结果表明,相比最新的深度学习水下图像增强方法,所提方法的结构相似性(SSIM)值提升了7.15%,峰值信噪比(PSNR)值提升了45.46%,该方法能够有效减少水下图像的雾度残留并增强图像细节。

Aiming at the problems of residual haze and blurred details in the current underwater image enhancement results,a Boosting underwater image enhancement method(BcGAN)based on conditional generative adversarial network(cGAN)was proposed.The SOS boosting strategy was introduced on the basis of the encoder-decoder structure to obtain a boosting generator and to realize the gradual refinement of image features.A dual discriminator structure was proposed to realize the discrimination of multi-scale input,and the joint training loss function of dual discriminators was constructed with WGAN-GP(Wasserstein GAN with gradient penalty)loss as the leading factor.Experimental results show that compared with the latest deep learning underwater image enhancement method,the structure similarity(SSIM)of the proposed method increases by 7.15%,and peak signal-to-noise ratio(PSNR)increases by 45.46%,which can effectively reduce the haze residue of underwater image and enhance the image details.