摘要
随着人工智能技术的快速发展,深度学习在图像处理领域取得了显著进展,特别是在目标检测和目标分割方面。传统的岩石识别方法受限于复杂的背景和岩石的多样性,无法满足实际需求。深度学习的快速发展为岩石智能识别提供了新的思路和技术支持。本研究旨在应用深度学习模型YOLOv8-seg于岩石智能识别任务中,评估其在目标检测和分割任务中的识别效果和稳定性,以期为地质勘探和地质资源管理提供技术支持。研究采用YOLOv8-seg模型,对包括玄武岩、花岗岩、大理岩、石英岩、煤炭、灰岩和砂岩在内的多种岩石类型进行训练,以优化模型的识别能力。该模型结合了目标检测和实例分割功能,并通过box_loss、seg_loss、cls_loss和dfl_loss等多种损失函数优化边界框预测、分割性能、类别识别准确性和回归精度。在目标分割任务中,YOLOv8-seg模型的precision(B)和recall(B)分别达到0.91284和0.93587,mAP50(B)和mAP50-95(B)分别为0.86666和0.83686;precision(M)和recall(M)分别为0.90394和0.93438,mAP50(M)和mAP50-95(M)分别为0.85931和0.81856,说明模型具备较高的分割精度和召回率。F1Score(B)和F1Score(M)在第551轮分别达至0.92421和0.91891,较初始值提升显著。测试集结果表明,模型在玄武岩、煤、灰岩等岩石类型的置信度均保持在90%以上,在岩石开采、煤炭运输等实际应用场景中的识别率保持在85%以上。YOLOv8-seg模型在岩石智能识别任务中表现出色,具有较高的精度、召回率和稳定性,适用于多种岩石分类和识别任务。结果表明,该模型在地质勘探和地质资源管理中具备广泛应用潜力,为岩石智能识别提供了一种可靠的解决方案。
With the rapid development of artificial intelligence technologies,significant advancements have been made for the application of deep learning in the field of image processing,particularly in the field of object detection and segmentation.Traditional rock recognition methods,limited by complex backgrounds and the diversity of rock types,can not meet the practical demand.The rapid growth of deep learning offers new approaches and technical supports for the intelligent rock recognition.This study aims to apply the deep learning model YOLOv8-seg into the intelligent rock recognition tasks,in order to evaluate the performance and stability of deep learning in object detection and segmentation,and to provide technical support for geological exploration and management of mineral resources.The YOLOv8-seg model was applied for the training to recognize various rock types including basalt,granite,marble,quartzite,coal,limestone,and sandstone,in order to optimize the intelligent rock recognition capability of the model.The model combines object detection and instance segmentation functions,and uses multiple loss functions,such as box_loss,seg_loss,cls_loss,and dfl loss,to optimize the bounding box prediction,segmentation performance,class recognition accuracy,and regression precision.For the object segmentation tasks by using the YOLOv8-seg model,precision(B)and recall(B)reached 0.91284 and 0.93587,mAP50(B)and mAP50-95(B)reached 0.86666 and 0.83686,precision(M)and recall(M)reached 0.90394 and 0.93438,mAP50(M)and mAP50-95(M)reached 0.85931 and 0.81856,respectively,indicating that high segmentation accuracies and recall rates can be achieved by applying the model.The F1 Score(B)and F1Score(M)respectively reached 0.92421 and 0.91891 in Round 551 of the training,showing significant improvements from their initial values.The results of the test set indicate that all average confidence levels of over 90%have been reached for the intelligent recognition of rock types such as basalt,coal,and limestone by applying the model.All average recognition rates of over 85%have been reached for the practical application scenarios like rock mining and coal transportation by using the model.Excellent performances,including high precision,recall,and stability,have been demonstrated in intelligent rock recognition tasks by applying the YOLOv8-seg model.Thus,the YOLOv8-seg model is suitable for completing various tasks of the classification and recognition of rocks.The research findings suggest that the model has broad potential of applications in geological exploration and management of mineral resources,and offers a reliable solution for intelligent rock recognition.
作者
何陆灏
周永章
张灿
HE Lu-hao;ZHOU Yong-zhang;ZHANG Can(Research Center for Global Environment and Earth Resources,Sun Yat-sen University,Guangzhou 510275,China;School of Earth Sciences and Engineering,Sun Yat-sen University,Zhuhai Guangdong 519085,China;Guangdong Provincial Key Laboratory of Geological Process and Mineral Resources Exploration,Zhuhai Guangdong 519085,China)
出处
《矿物岩石地球化学通报》
北大核心
2025年第3期525-541,共17页
Bulletin of Mineralogy, Petrology and Geochemistry
基金
国家重点研发计划项目(2022YFF0801201)
国家自然科学基金资助项目(U1911202)
广东省重点领域研发计划项目(2020B1111370001)。
关键词
岩石识别
深度学习
机器学习
卷积神经网络
目标分割
图像识别
rock recognition
deep learning
machine learning
convolutional neural network
object segmentation
image recognition
