基于卷积网络的边缘保持滤波方法

Edge-preserving Filtering Method Based on Convolutional Neural Networks

边缘保持滤波是计算机视觉、图像处理领域的重要基础理论研究,作为图像预处理操作对后续的处理结果有着重要影响。区别于传统滤波方法,边缘保持滤波方法不仅注重图像的平滑处理,还注重保持边缘细节。卷积神经网络在很多领域得到了应用,并取得显著的成果。本文将卷积神经网络引入边缘保持滤波,利用卷积神经网络的良好扩展性和灵活性来构建深度卷积神经网络模型(Deep Convolutional Neural Network,DCNN),通过3种类型的网络堆叠层,采用反向传播迭代更新网络参数,训练残差图像,实现基于DCNN的边缘保持滤波方法;还构建了基于梯度域的卷积神经网络模型(Gradient CNN,GCNN),对彩色图像的梯度信息进行学习,通过三层卷积对梯度图进行边缘保持平滑操作,得到边缘保持平滑梯度图,进而利用输入图像引导平滑梯度图进行彩色重建,得到彩色滤波图像。最后通过实验与常见的边缘保持滤波方法进行主观和客观评价对比。DCNN不仅在视觉上达到了其他滤波的效果,同时在处理时间上也存在较大优势,表明DCNN可以通过大量的数据训练有效地拟合出多种边缘保持滤波算法。与其他边缘保持滤波结果相比,GCNN在视觉上可以保持颜色风格与输入图像整体一致,而且图像相似度评价指标也更好,表明GCNN解决了部分滤波处理出现颜色偏差、梯度反转等问题,而且提高了处理效率。

Edge-preserving filtering is a significant basic theory research in the fields of computer vision and image processing.As subsequent operation of pre-processing,edge-preserving filtering has great influence on final results of image processing.Different with traditional filtering,edge-preserving filtering focuses not only on smooth,but also on image edge details.Convolutional neural networks (CNNs) have been applied into a variety of research fields with great success.In this paper,CNN was introduced into edge-preserving filtering.Taking advantages of CNN’s excellent extensibility and flexibility,this paper constructed a deep convolutional neural network (DCNN).With three types of cascading network layers,DCNN iteratively updates its parameters by back propagation,produces a residual image and realizes a DCNN-based edge-preserving filtering.Besides,a gradient CNN model (GCNN) was constructed.The gradient of color images is learnt,edge-preserving smoothing operation is conducted for gradient images by three layers of convolution,and edge-preserving filtering gradient images are obtained.Subsequently,the input image is used to guide the filtering gradient image for reconstructing and obtaining color filtering image.Finally,experiments were made to evaluate the proposed methods and the proposed methods were compared with popular edge-preserving filtering methods subjectively and objectively.DCNN not only achieves the same visual effects as other methods,but also has big advantages in processing time,which demonstrates that DCNN can effectively and efficiently imitate various filtering methods through training on large amount of data.For GCNN,in terms of visual effects,its output conforms to the input in the color style globally.In terms of image similarity evaluation,it also outperforms other methods.This verifies that GCNN can address the problems of color shift and gradient inversion,as well as improve the filtering efficiency.