基于多尺度LBP算子的大理石纹理分割算法  

Marble texture segmentation algorithm based on multi-scale LBP operator

作　　者：邹智伟 张战成 姚浩男 徐少康 ZOU Zhiwei;ZHANG Zhancheng;YAO Haonan;XU Shaokang(School of Electronic&Information Engineering,SUST,Suzhou 215009,China)

机构地区：[1]苏州科技大学电子与信息工程学院,江苏苏州215009

出　　处：《苏州科技大学学报（自然科学版）》2023年第3期71-76,共6页Journal of Suzhou University of Science and Technology（Natural Science Edition）

基　　金：国家自然科学基金面上项目(61772237)。

摘　　要：纹理分割算法是仿大理石石材生产的重要步骤。大理石图像中各种纹理特征交杂在一起,当前主流纹理分割算法无法将其清楚、连续地分割。受空洞卷积(Dilated Convolutions)的启发,提出一种多尺度LBP算子(MLBP),该算子将边缘和内部像素点组合排列,生成多个尺度算子,增大了特征提取的感受野,通过多个子尺度算子的融合能捕获到纹理的变化,可以适应大理石图像纹理风格多变的特点,增强特征提取的泛化能力。基于GMM聚类算法,在真实大理石数据集上验证了纹理分割效果,像素精度(PA)为93.2%,平均像素精度(MPA)为89.3%,平均交并比(MIoU)为71.2%,与FCN、K-means和FCM等聚类算对比实验显示,MLBP特征算子优于传统的LBP特征算子。Texture segmentation algorithm is an important step in the production of the ceramic tile which imitates the style of marble.Various texture features in marble images are intertwined,and existing mainstream texture segmentation algorithms cannot segment them clearly and continuously.Inspired by Dilated Convolutions,a multi-scale LBP operator(MLBP)is proposed,which combines the edge and internal pixels to generate multiple scale operators.Thus the receptive field of feature extraction is increased.Moreover,the texture variation can be captured through the fusion of multiple sub-scale operators,accommodating to the polytropic texture style of marble images and enhancing the generalization ability of feature extraction.Based on the GMM clustering algorithm,the texture segmentation effect is verified on the real marble dataset.The pixel accuracy(PA)is 93.2%,the average pixel accuracy(MPA)is 89.3%,and the average intersection ratio(MIoU)is 71.2%.Compared with FCN,K-means and FCM clustering on raw image and LBP,the proposed MLBP feature operator is better than the traditional LBP feature operator.

关 键 词：纹理分割 LBP算子 混合高斯 大理石纹理 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]