Abstract:

Aiming at the limitation of the current dehazing technology and the slow processing speed, a lightweight dehazing algorithm based on multi-scale convolutional network is proposed. The algorithm has two parts：dehazing and repairing. The dehazing module accepts the foggy image. After feature extraction and fusion, the transmittance map and the atmospheric light value are uniformly learned. And the deformed atmospheric physical scattering algorithm is used to invert the fog-free image. But the fog-free image still has such problems as the dark hue and the unclear details. The repairing module repairs the image from the dehazing module by limiting the contrast histogram equalization（CLAHE） to improve the image contrast and the robustness of the algorithm. In the experiments, the dehazing task is combined with the target detection task to verify the validity of the algorithm.

Key words: dehazing algorithm, convolutional network, computer vision, deep learning

摘要：

针对目前图像去雾技术存在的使用场景有限、处理速度慢等问题，提出一种基于多尺度卷积网络的快速去雾算法。算法由去雾和修复两部分组成。去雾模块首先将有雾图像输入，经过特征提取和融合，然后通过变形后的大气物理散射算法对透射率图和大气光值统一学习，并演出去雾图像。去雾后的图像仍存在色调偏暗、细节不清晰的问题。修复模块利用对比度受限自适应直方图均衡方法对去雾图像进行修复，提升图像的对比度和算法的鲁棒性。通过去雾任务与目标检测任务相结合的测试实验进一步验证了算法的有效性。

关键词: 去雾算法, 卷积网络, 计算机视觉, 深度学习