基于复杂纹理特征融合的材料图像分割方法
Material image segmentation method based on complex texture feature fusion
韩越兴 1杨珅 2陈侨川 2王冰2
作者信息
- 1. 上海大学计算机工程与科学学院,上海 200444;之江实验室,浙江杭州 311100
- 2. 上海大学计算机工程与科学学院,上海 200444
- 折叠
摘要
为解决材料图像分割中存在小样本、纹理复杂和数据分布不平衡的问题,抓住材料图像同相像素具有高度相似性的特性,提出一种基于复杂纹理特征融合的材料图像分割方法.在编码阶段,使用全卷积神经网络(FCN)作为基础网络,VGG16作为骨干网络;将改进的FCN的每层的特征图放入设计的级联的特征融合模块(CFF block),融合高低层语义信息;将融合的特征图放入多尺度学习模块(multi-scale block)进一步提取纹理特征.在解码阶段,对特征图施加注意力机制(Attention block),保留关键的特征图;针对材料图像中数据不平衡问题,采用并改进Dice损失,优化分割结果.通过对比实验和消融实验验证该方法的mIoU在多个数据集上均优于经典的深度学习方法.
Abstract
To solve the problems of small sample,complex texture and unbalanced data distribution in material image segmenta-tion,the characteristics of high similarity of pixels in the same phase were grasped,a material image segmentation method based on complex texture feature fusion was proposed.In the coding phase,the convolutional neural network network(FCN)was used as the base network and VGG16 as the backbone network,the feature map of each layer of the improved FCN was put into the designed cascade feature fusion module(CFF block),and the fused feature map was put into the multi-scale block to extract tex-ture features.In the decoding stage,the Attention block was applied to the feature map,and the key feature map was retained.To solve the problem of unbalanced data distribution in material images,Dice loss was adopted and improved to optimize the seg-mentation results.Results of comparison experiment and ablation experiment show that the mIoU of the proposed method is bet-ter than that of classical deep learning methods in many data sets.
关键词
材料图像分割/全卷积神经网络/特征融合/Dice损失/交叉熵损失/注意力机制/小样本Key words
material image segmentation/FCN/feature fusion/Dice loss/cross-entropy loss/Attention mechanism/small sample引用本文复制引用
基金项目
国家重点研发计划基金项目(2018YFB0704400)
国家自然科学基金面上基金项目(52273228)
上海市自然科学基金项目(20ZR1419000)
之江实验室科研攻关基金项目(2021PEOAC02)
出版年
2024
NETL
NSTL
万方数据