摘要
为解决基于视觉的船舶智能感知系统在海雾悬浮颗粒的散射作用下,图像信息可见度与对比度明显下降,目标检测、目标跟踪和语义分割等高层视觉任务受到严重影响,导致无法检测到目标或跟踪目标丢失的问题,创建了Maritime Haze数据集,并结合物理模型和深度学习方法,提出了融合暗通道先验知识的循环生成对抗网络图像去雾模型;利用暗通道先验去雾算法将图像分解输出透射图和去雾图像,再通过循环生成对抗网络的生成器和判别器对暗通道先验输出进行处理和判断,从而生成更好的去雾图像。实验表明,该模型峰值信噪比(PSNR)和结构相似性(SSMI)的评价指标分别达24.73和0.943,优于PSD的23.34(PSNR)和0.921(SSMI)以及CycleGAN的19.19(PSNR)和0.584(SSMI),并在视觉效果上展示出最佳的清晰度和色彩真实性,在航海图像去雾领域领先其他方法。
In order to solve the problem of ship intelligent perception system based on the vision, under the scattering effect of sea fog suspended particles, the visibility and contrast of image information has significantly reduced, the high-level visual tasks such as target detection, target tracking and semantic segmentation were seriously affected, resulting in the inability to detect targets or the loss of tracking targets, the Maritime Haze data set was created, and by combining with physical models and depth learning methods, a cycle generation adversarial network image dehazing model based on prior knowledge of dark channel was proposed. A dark channel prior dehazing algorithm was used to decompose an image into transmission map and dehazing image. Then, the generator and discriminator of a cycle generative adversarial network were used to process and judge the dark channel prior output to generate a better dehazing image. The experiments show that the evaluation indexes of peak signal to noise ratio(PSNR) and structural similarity(SSMI) of the model reaches 24.73 and 0.943 respectively, which are superior to 23.34(PSNR) and 0.921(SSMI) of PSD and 19.19(PSNR) and 0.584(SSMI) of CycleGAN with excellent clarity and color authenticity in visual effects, and leading other methods in marine image dehazing.
作者
张成
潘明阳
高翊然
王泾阳
张若澜
李邵喜
ZHANG Cheng;PAN Ming-yang;GAO Yi-ran;WANG Jing-yang;ZHANG Ruo-lan;LI Shao-xi(Navigation College,Dalian Maritime University,Dalian 116026,China)
出处
《大连海事大学学报》
CAS
CSCD
北大核心
2022年第4期84-93,共10页
Journal of Dalian Maritime University
基金
广西壮族自治区科技厅重点研发项目(2021AB07045)。
