形态学与RCF相结合的唐卡图像边缘检测算法

唐卡图像的内容丰富,纹理信息复杂。边缘检测在唐卡图像分析研究中具有非常重要的意义,因为唐卡图像轮廓含有大量的图像数据信息。数学形态学方法提取的边缘光滑连续,但是对复杂的边缘检测时会存在模糊不清晰的现象[1]。卷积神经网络(CNN)可以提取很多高层的、多尺度的信息[2]。为此提出的边缘检测方法,用优化的数学形态学算法提取原图像边缘;利用训练的RCF网络模型[3]提取原图像的边缘。根据小波变换的分解与重构原理将以上方法得出的图像边缘融合,从而得到更加完整光滑的图像边缘。实验表明,融合后的图像边缘更加清晰连续,轮廓信息更符合人类的视觉认知,去掉了无效的细节纹理,更有利于唐卡图像后续研究。

THANG KA IMAGE EDGE DETECTION ALGORITHM BASED ON MORPHOLOGY AND RCF

Thang ka image is rich in content and complex in texture information.Edge detection is very important in the analysis of Thang ka image,because Thang ka image contour contains a lot of image data information.The edges extracted by mathematical morphology method are smooth and continuous,but there is ambiguity in complex edge detection.Convolutional neural network(CNN)can extract a lot of high-level and multi-scale information.The proposed edge detection method used the optimized mathematical morphology algorithm to extract the original image edge.Then,the training RCF network model was used to extract the edges of the original image.According to the decomposition and reconstruction principle of wavelet transform,we obtained the image edges by the above method to fuse,getting more complete and smooth image edges.Experiments show that the fused image edge is clearer and more continuous,the contour information is more in line with human visual cognition,and the invalid detail texture is removed.It is more conducive to the follow-up study of Thang ka image.