基于对比约束的轻量图像去雾网络

Contrast Constraint-Based Lightweight Image Dehazing Network

近年来基于卷积神经网络(CNN)的方法广泛用于图像去雾领域,并获得明显的性能提升。虽然基于CNN的方法取得巨大成功,但现有去雾方法需要大量运算资源,难以部署到边缘设备或嵌入式设备中。为了解决这个问题,本文提出了一种基于对比约束的轻量化图像去雾网络。整体结构采用U-Net架构,降维雾霾图像以降低运算资源消耗。采用了一种双蝶形栅格模块作为基础模块,栅格模块灵活结合了残差连接和注意力机制,能学习更具有判别性的特征表示。该网络以端到端的形式学习图像去雾映射,避免中间估计带来的误差。实验结果表明该方法在客观指标和主观感知质量上均优于其他对比方法。

CNN-based methods have been widely used for image dehazing and achieved significant performance improvements in recent years.Although CNN-based methods have succeeded,they are difficult to deploy into edge devices or embedded devices due to the significant computational resources required for exiting methods.This paper proposes a lightweight image dehazing network based on contrast constraints to solve this problem.The overall structure adopts the U-Net architecture to reduce the dimensionality of haze images and reduce computing resource consumption.This paper adopts the lattice block(LB)as the basic module of this paper,which contains two such butterfly structures with residual learning.The LB flexibly combines residual connection and attention mechanism,which can learn more discriminative feature representation.The network learns the image dehazing map end-to-end,avoiding errors caused by intermediate estimates.The experimental results show that the method outperforms other comparison methods in objective indicators and subjective perception quality.