摘要
在地质数据不断增加的背景下,深度学习技术已在矿物识别领域取得巨大进展和广泛应用。然而,现有的多尺度卷积神经网络模型存在速度较慢、鲁棒性不足、准确性有待提高等问题,这限制了其在复杂地质环境中的矿物识别应用。针对这些问题,本次研究旨在优化多尺度卷积神经网络模型,以在矿物识别的准确性、速度和鲁棒性方面取得更好的效果。本次研究提出了一个改进的卷积神经网络模型VGG-16,该模型通过采用深度可分离卷积、动态卷积层、特征金字塔网络、通道注意力机制和局部感知域卷积层5种改进方法优化模型,本次实验基于地质数据集验证了VGG-16模型的改进效果,使用包含大量高质量地质样本的数据集进行训练和测试,模型的准确性得到了明显提高。实验结果显示,改进模型在矿物识别方面取得了显著的效果提升,对7种矿物增强识别后识别准确率达到了92.2%,尤其在处理复杂地质环境的矿物识别任务时,在处理颜色复杂、纹路难以辨别的岩石图像时,对于不同尺寸和旋转角度的岩石也能更好地进行识别,表现出了优于传统多尺度卷积神经网络模型的性能,该优化模型在识别准确率、训练速度和鲁棒性方面均得到了改善,为矿物识别领域的研究和应用提供了有力的支持。
In the context of increasing geological data,deep learning technology has made great progress and wide application in the field of mineral identification.However,the existing multi-scale convolutional neural network model has such problems as slow speed,insufficient robustness,and low accuracy,which limit its application in complex geological environment.In this paper,an improved convolutional neural network model VGG-16 is proposed to optimize the multi-scale convolutional neural network model by using five improved methods of deep separable convolutional layer,including dynamic convolutional layer,feature pyramid network,channel attention mechanism and local perception domain convolutional layer,so as to achieve better results in the accuracy,speed and robustness in mineral identification.The experiment introduced in this paper verifies the improvement effect of VGG-16 model based on geological data sets containing a large number of high-quality geological samples and this model is used for training,testing and improving the recognition accuracy.The results show that the improved model has achieved significant improvement in mineral recognition,with a recognition accuracy of 92.2%after enhanced recognition of 7 minerals.This improved model can well process and identify rocks of different sizes and rotation angles when identifying minerals in complex geological environments,such as rock images with complex colors and difficult line patterns.Hence,its performance is better than that of traditional multi-scale convolutional neural network models,and the optimized model has been improved in the aspects of recognition accuracy,training speed and robustness,thus providing strong support for the research and application of mineral recognition.
作者
柳博文
刘星
LIU Bowen;LIU Xing(Anhui University of Science and Technology,Huainan,Anhui 232000,China)
出处
《矿物岩石》
CAS
CSCD
北大核心
2023年第3期10-19,共10页
Mineralogy and Petrology
基金
国家自然科学基金面上项目(批准号:51974007)。
关键词
矿物识别
VGG-16
深度可分离卷积
动态卷积层
特征金字塔网络
通道注意力机制
局部感知域卷积层
mineral recognition
VGG-16
deep separable convolution
dynamic convolution layer
crossscale feature fusion
separable attention mechanism
local perceptual domain convolution layer
