光学 精密工程, 2021, 29 (8): 1931, 网络出版: 2021-10-25
分离特征和协同网络下的端到端图像去雾
End-to-end image dehazing under separated features and collaborative network
图像去雾 卷积神经网络 大气散射模型 特征分离 图像复原 image dehazing convolutional neural network atmospheric scattering model feature separation image restoration
摘要
针对去雾领域中传统方法受手动设置特征的限制,以及现有网络去雾不彻底和细节保持不佳等问题,提出一种分离特征和协同网络下的端到端图像去雾模型。首先对传统的大气散射模型进行变形,分离出乘性特征和加性特征。其次,根据两个特征对最终去雾结果的影响程度,设计基于乘性特征和加性特征提取框架并行驱动的去雾网络。其中,乘性特征提取网络充分考虑了不同深度的空间信息及细节特征,通过各层之间密集级联达到特征重用和信息补偿的目的,以获取精密丰富的目标特征。另外,利用残差跨连结构搭建加性特征提取网络,用于训练偏置加性特征。最后,将分离特征代入复原模型得到无雾图像。实验表明:所提网络去雾效果显著,复原图像颜色自然,细节保持良好且各项指标占优。
Abstract
Hand-designed features limit the performance of traditional dehazing methods, and existing networks encounter problems such as incomplete dehazing and significant loss of detail. Therefore, a model of end-to-end dehazing with separated features and a collaborative network is proposed herein. First, the traditional atmospheric scattering model is transformed to separate the multiplicative and additive features. Second, according to the influence of two features on the final dehazing result, a parallelly driven dehazing architecture is designed based on multiplicative and additive feature extraction frameworks. Further, spatial information and detailed features of different depths are fully considered in the multiplicative feature extraction network, and feature reuse and information compensation are realized using dense cascading to obtain precise and rich target features. Additionally, an additive feature extraction network is built to acquire biased and additive features according to the residual cross-connection structure. Finally, separated features are substituted into the restoration model to obtain a haze-free image. Experiment results show that the proposed network offers a significant dehazing effect, natural colors of restored images, outstanding detail retention, and superior scores of various metrics.
杨燕, 梁小珍, 张金龙. 分离特征和协同网络下的端到端图像去雾[J]. 光学 精密工程, 2021, 29(8): 1931. Yan YANG, Xiao-zhen LIANG, Jin-long ZHANG. End-to-end image dehazing under separated features and collaborative network[J]. Optics and Precision Engineering, 2021, 29(8): 1931.