基于YOLOv8-seg的岩心CT图像颗粒目标提取算法  

Particle target extraction algorithm for rock core CT image based on YOLOv8-seg

作　　者：郭俊青 何小海[1] 滕奇志[1] 吕朝阳 GUO Jun-Qing;HE Xiao-Hai;TENG Qi-Zhi;LÜ Zhao-Yang(College of Electronics and Information Engineering,Sichuan University,Chengdu 610065,China;Chengdu Xitu Technology Co.,Ltd,Chengdu 610065,China)

机构地区：[1]四川大学电子信息学院,成都610065 [2]成都西图科技有限公司,成都610065

出　　处：《四川大学学报(自然科学版)》2025年第1期116-125,共10页Journal of Sichuan University(Natural Science Edition)

基　　金：国家自然科学基金(62071315)。

摘　　要：由于岩心CT图像噪点多、亮度不均、对比度低,颗粒之间高度粘连,导致难以直接对颗粒进行提取分析.针对这一问题,本文提出一种基于YOLOv8-seg的岩心CT图像颗粒目标提取改进算法.基于YOLOv8-seg的骨干架构,结合多头自注意力(Multi-Head Self-Attention,简称MHSA)的思想,将CBAM和BoTNet两部分融合为一个全新的多头处理模块CBoTNet,实现通道和空间全局多头注意力机制,加强了特征提取的能力;使用Wise-IoU替换边界框回归损失函数CIoU,有效减少了由低质量样本产生的有害梯度,进一步提高了模型精度.在自制岩心CT序列颗粒图像数据集中,相较于原始YOLOv8-seg算法,mAP50精度提高了1.30%,mAP50:95提高了7.81%.实验结果证明,与其他实例分割网络相比,该算法能较为准确地提取岩心颗粒,并解决颗粒之间的粘连,具有较好的精确度和稳定性.Due to the high noise,uneven brightness,low contrast,and high adhesion between particles in core CT images,it is difficult to directly extract and analyze particles.To address this issue,an improved algorithm based on YOLOv8-seg has been proposed for particle target extraction from core CT images.The algorithm is developed by integrating the ideas of Multi-Head Self-Attention(MHSA),CBAM,and BoTNet into a newly designed multi-head processing module called CBoTNet,in which channel and spatial global multi-head attention mechanisms are implemented to enhance feature extraction.Moreover,the model accuracy is further enhanced through the substitution of the boundary loss CIoU with Wise-IoU,and harmful gradients generated by low-quality samples are effectively reduced.On a self-constructed core CT sequence particle image dataset,mAP50 accuracy is improved by 1.30%,and mAP50:95 accuracy is increased by 7.81%when compared to the original YOLOv8-seg algorithm.Experimental results demonstrate that,compared to other instance segmentation networks,core particles can be accurately extracted,adhesion between particles is resolved,and superior accuracy and stability are achieved.

关 键 词：YOLOv8-seg 岩心CT图像 颗粒目标提取 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]