Abstract: 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.

Key words: Convolutional neural networks, Edge-preserving, Filtering, Smoothing