基于多尺度生成对抗网络的水下图像增强

Underwater Image Enhancement Based on Multiscale Generative Adversarial Network

针对水体对光的吸收和散射导致的水下图像细节模糊和颜色失真等问题,提出了一种基于多尺度生成对抗网络的水下图像增强算法。该算法用对抗网络作为基础框架,结合残差连接和密集连接加强水下图像特征的传播。首先,通过两个并行支路提取退化图像不同空间的视觉信息,并在每个支路加入残差密集块,以学习更深层次的特征。然后,将两个支路提取的特征进行融合,经过重建模块恢复图像的细节信息。最后,构建多个损失函数,反复训练对抗网络,获得增强的水下图像。实验结果表明,本算法增强的水下图像色彩鲜明且去雾效果较好,水下彩色图像质量均值比原始图像高0.1887,加速稳健特征的匹配点数比水下残差网络算法多17个。

To address problems associated with capturing underwater images,i.e.,blur details and color distortion caused by the absorption and scattering of light,an underwater image enhancement algorithm based on multiscale generative adversarial network is proposed.This algorithm uses an adversarial network as the basic framework,combining residual connections and dense connections to strengthen the propagation of underwater image features.First,the visual information in different spaces of a degraded image is extracted through two parallel branches,and a dense residual block is added to each branch to learn deeper features.Then,the features extracted from the two branches are fused and the detailed information of the image is restored through a reconstruction module.Finally,multiple loss functions are constructed and the adversarial network is repeatedly trained to obtain enhanced underwater images.The experimental results demonstrate that an underwater image enhanced using the algorithm has brighter colors and better dehazing effect.Compared with the original image,the average quality of the underwater color image is increased by 0.1887;compared with the underwater residual network algorithm,the number of matching points of the speeded up robust features is increased by 17.