跨层双线性融合改进EfficientNet-B5的笔石图像分类方法  

Graptolite image classification using the EfficientNet-B5 improved by cross-layer bilinear fusion

作　　者：梁源潼 李朝锋 刘洪林[2,3,4] 张介辉 计玉冰 李晓波 LIANG Yuantong;LI Chaofeng;LIU Honglin;ZHANG Jiehui;JI Yubing;LI Xiaobo(Institute of Logistics Science and Engineering,Shanghai Maritime University,Shanghai 201306,China;China Petroleum Exploration and Development Research Institute,Beijing 100083,China;PetroChina Unconventional Oil and Gas Key Laboratory,Beijing 100083,China;National Energy Shale Gas Research and Development(Experiment)Center,Langfang 065007,China;Zhejiang Oilfield,Petrochina,Hangzhou 211762,China)

机构地区：[1]上海海事大学物流科学与工程研究院,上海201306 [2]中国石油勘探开发研究院,北京100083 [3]中国石油非常规油气重点实验室,北京100083 [4]国家能源页岩气研发(实验)中心,河北廊坊065007 [5]中国石油浙江油田公司,浙江杭州211762

出　　处：《应用科技》2022年第5期15-23,共9页Applied Science and Technology

摘　　要：笔石化石的分布为页岩地层的精细划分提供了重要依据。针对笔石种类多、种间特征差别小导致人工鉴定难度大的问题,提出一种跨层双线性融合改进EfficientNet-B5的笔石图像分类方法,跨层特征融合模块有效地增强了模型的特征提取能力。首先将图像输入EfficientNet-B5模型,依据不同层次特征图的特点,分别将EfficientNet-B5第2、第5、第7个移动翻转瓶颈卷积块输出的特征图选作低层、高层和全局特征信息;采用跨层双线性融合方法将多层次特征融合;最后将融合特征送入Softmax分类器,输出图像的预测标签。基于国内常见笔石,构建了51种笔石的图像数据集并进行实验,本文方法的分类精度达到94.03%,较其他图像分类方法具有一定优势。The distribution of graptolite fossils provides an important basis for the fine division of shale strata.However,there are many kinds of graptolites and little difference in characteristics between species,it is hard to identify them manually.A method for graptolite classification based on cross-layer bilinear fusion and EfficientNet-B5 is proposed.Firstly,an image is input into the model.According to the characteristics of different feature maps,the feature maps output by the 2nd,5th and 7th MBConv blocks of the model are selected as low-level,high-level and global features respectively.The multi-level features are fused by cross-layer bilinear fusion method.Finally,the fused features are sent to the classifier to output the prediction label.Based on the common graptolites in China,the image data sets of 51 kinds of graptolite are constructed and tested.The accuracy of this method reaches 94.03%.Compared with other classification methods,our method shows some advantages.

关 键 词：图像处理 图像分类 笔石化石 卷积神经网络 跨层双线性融合 特征图 可视化 数据扩充 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]