基于深度学习的硬质合金图像分割算法  

Deep learning-based image segmentation algorithm for tungsten carbide

作　　者：阮鹏 何小海[1] 滕奇志[1] RUAN Peng;HE Xiao-Hai;TENG Qi-Zhi(School of Electronics and Information Engineering,Sichuan University,Chengdu 610065,China)

机构地区：[1]四川大学电子信息学院,成都610065

出　　处：《四川大学学报(自然科学版)》2025年第2期388-398,共11页Journal of Sichuan University(Natural Science Edition)

基　　金：国家自然科学基金(62071315)。

摘　　要：为了提取硬质合金显微图像的碳化钨晶粒区域,本文提出一种基于多层级注意力融合机制的深度学习网络分割算法.基于UNet网络模型,将深层下采样结果通过多层级注意力融合模块,增强下采样特征图中碳化钨晶粒的特征信息,学习每个通道的权重,通过该权重突出晶粒的特征信息,提高模型对重要特征通道的关注度.再与上采样结果进行跳跃连接,加强特征图对位置间的关联,提高模型对输入数据的表达能力.最后在不同类型的硬质合金测试集上进行了对比和消融实验.结果表明,本文提出的方法对分割硬质合金显微图像的有效性.A deep learning network segmentation algorithm based on a multi-level attention fusion mechanism is proposed to extract the tungsten carbide grain regions in micrographs of hard alloy.Utilizing the UNet network model,a multi-level attention fusion module is incorporated within the skip connections between deep down-sampling and up-sampling layers to enhance the feature representation of tungsten carbide grains in down-sampled feature maps.By learning the weights for each channel,this mechanism accentuates the grain features,thereby enhancing the model's focus on crucial feature channels.This strengthens the correlation between features and improves the model's capability to represent input data.Comparative and ablation experiments are conducted on various types of hard alloy test sets,demonstrating the effectiveness of the proposed approach in segmenting micrographs of hard alloys.

关 键 词：硬质合金显微图像 图像分割 对比度 深度学习 自注意力机制 

分 类 号：TP341[自动化与计算机技术—计算机系统结构]