结合梯度指导和局部增强Transformer的图像去模糊网络

Image Deblurring Model Combining Gradient Guidance and Local Enhancement Transformer

模糊图像不仅影响人类感知还会影响后续计算机视觉任务的性能,例如自动驾驶系统和户外监控系统中的视觉算法.针对以往基于深度学习的去模糊方法感受野较小,不能动态适应输入内容和重建图像细节信息困难等问题,提出了一种基于Transformer的图像去模糊网络.网络包含两个分支:图像内容分支和梯度分支,每条分支均以具有窗口机制的Transformer作为主干,通过梯度分支的信息指导图像去模糊重建,能够更好地恢复图像的边缘和纹理.同时,为了充分利用图像的内容信息和梯度信息,本文还设计了一个交互式融合模块来有效融合特征信息.此外,本文通过在Transformer块的自注意力机制和前馈网络中引入卷积来解决Transformer对局部信息建模不足的问题.在合成数据集和真实数据集上的大量实验结果表明,提出的算法能有效去除复杂模糊并且恢复清晰的细节,在定量指标和视觉效果上均优于目前的主流去模糊算法.

Blurred images not only affect human perception but also the performance of subsequent computer vision tasks,such as vision algorithms in autonomous driving systems and outdoor surveillance systems.To address the issues that previous deep learning-based deblurring methods with small receptive fields,cannot dynamically adapt to the input content and have difficulties in reconstructing image detail information,a Transformer-based image deblurring network is proposed.The network contains two branches:the image content branch and the gradient branch,each branch takes windows-based Transformer as the backbone,and guides the image deblurring reconstruction by the information of the gradient branch,which can better recover the edges and textures of the image.Meanwhile,in order to make full use of the content details and gradient information of the image,this paper also designs an interactive fusion module to effectively fuse the feature information.In addition,in order to solve the problem of Transformer′s insufficient ability to model local information,this paper introduces convolution in the self-attention mechanism and feedforward network of Transformer block.Experimental results on synthetic and real datasets demonstrate that the proposed algorithm can effectively remove complex blur and recover clear details,and outperforms current mainstream deblurring algorithms in terms of quantitative metrics and visual effects.