基于动态卷积核的自适应图像去雾算法

Adaptive Image Dehazing Algorithm Based on Dynamic Convolution Kernels

现有图像去雾方法普遍存在去雾不彻底、容易出现颜色失真等问题,基于传统深度学习模型的图像去雾方法多采用静态推理模式,在该模式下,模型对不同样本会采用同样的、固定的参数设置,从而抑制了模型的表达能力,影响图像的去雾效果。针对以上问题,文中提出了一种基于动态卷积核的自适应图像去雾算法,该算法包括编码网络、自适应特征增强网络和解码网络3个部分。文中采用动态卷积、密集残差、注意力机制设计了自适应特征增强网络,该网络主要包括动态残差组件和动态跨层特征融合组件。动态残差组件由动态密集残差模块、一个卷积层和双注意力模块构成,其中动态密集残差模块将动态卷积引入密集残差模块,同时设计了一个基于注意力的权重动态聚合子网络,动态地生成卷积核参数以达到样本自适应的目的,在减少信息丢失的同时增强了模型的表达能力;双注意力模块结合通道注意力和像素注意力,使模型更加关注图像通道之间的差异性以及雾霾分布不均匀的区域。动态跨层特征融合组件通过动态融合不同阶段的特征,来学习丰富的上下文信息,防止网络深层计算时遗忘网络的早期特征,同时极大地丰富了特征表示,有利于模型对无雾图像细节信息的恢复。在合成数据集和真实数据集上进行了大量实验,结果表明,所提方法不仅取得了较好的客观评价分数,而且重建了主观效果较好的去雾图像,超越了对比方法的性能。

Existing image dehazing methods generally have problems such as incomplete dehazing and color distortion.Image dehazing methods based on traditional deep learning models mostly use static inference during testing,which use the same and fixed parameters for different samples,thereby inhibiting the expressive ability of the model and decreasing the dehazing performance.Aiming at the above problems,this paper proposes an adaptive image dehazing algorithm based on dynamic convolution kernel.The proposed model includes three parts:encoding network,adaptive feature enhancement network and decoding network.This paper combines dynamic convolutions,dense residual connections,and attention mechanism to complete the adaptive feature enhancement network,which mainly includes dynamic residual components and dynamic skip-connected feature fusion components.The dynamic residual component is composed of a dynamic residual dense block,a convolutional layer and a dual attention mo-dule.The dynamic residual dense block introduces dynamic convolutions into the residual dense block,and an attention-based weight dynamic aggregator is designed at the same time,which dynamically generates adaptive convolution kernel parameters.The dynamic convolutions have reduced the loss of information and enhanced the expressive ability of the model.The dual attention module combines channel attention and pixel attention to make the model pay more attention to the differences between image channels and areas with uneven distribution of haze.The dynamic skip-connected feature fusion component learns rich contextual information by dynamically fusing the features of different stages via skip-connections,preventing the early features of the network from being forgotten when the information flows into deeper layers.Meanwhile,the feature representations are greatly enriched,which benefits the restorations of the details for fog-free images.Extensive experiments on synthetic datasets and real datasets show that our method not only achieves better objective evaluation scores,but also reconstructs dehazing images with better visual effects,surpassing the performance of compared methods.