彭伟航, 白林, 商世为, 唐小洁, 张哲远. 基于改进InceptionV3模型的常见矿物智能识别[J]. 地质通报, 2019, 38(12): 2059-2066.
    引用本文: 彭伟航, 白林, 商世为, 唐小洁, 张哲远. 基于改进InceptionV3模型的常见矿物智能识别[J]. 地质通报, 2019, 38(12): 2059-2066.
    PENG Weihang, BAI Lin, SHANG Shiwei, TANG Xiaojie, ZHANG Zheyuan. Common mineral intelligent recognition based on improved InceptionV3[J]. Geological Bulletin of China, 2019, 38(12): 2059-2066.
    Citation: PENG Weihang, BAI Lin, SHANG Shiwei, TANG Xiaojie, ZHANG Zheyuan. Common mineral intelligent recognition based on improved InceptionV3[J]. Geological Bulletin of China, 2019, 38(12): 2059-2066.

    基于改进InceptionV3模型的常见矿物智能识别

    Common mineral intelligent recognition based on improved InceptionV3

    • 摘要: 以常见的16类矿物作为研究对象,收集每一类矿物约1000张图像,按比例划分为训练集、验证集和测试集,通过图像随机选取增加数据的多样性,建立矿物识别InceptionV3模型,训练7万次在测试集上获得81%的识别正确率。通过对损失函数的改进,引入Center Loss损失函数,训练40万次识别准确率提高到86%。对分类的混淆矩阵分析发现,孔雀石等外观特征明显的矿物识别正确率很高,而闪锌矿等与其他矿物容易混淆导致正确率较低。从特征图分析看出,模型很好地提取了孔雀石的放射状特征,矿物图像特征向量聚集程度很高,也说明了模型的可靠性。

       

      Abstract: To study 16 kinds of common minerals, the authors collected 1000 images for each type, and then divided them into training set, validation set and test set. Before putting the images into the model, the authors selected a random area of each image for data augmentation. After training the InceptionV3 model with 70000 steps, the authors obtained an 81% accuracy in the test set. Through improving the loss function and introducing the Center Loss, the authors raised the accuracy to 86% after training 400000 steps. The obfuscation matrix shows that, the recognition accuracies for the minerals with obvious appearance characteristics such as malachite are higher while those for other minerals like sphalerite are less due to the obfuscation with other minerals. The analysis of the feature map shows that the model extracts the radial feature of malachite perfectly, and the feature vector of mineral image aggregate is in a high degree, which also can prove the reliability of the model.

       

    /

    返回文章
    返回