Fusulinid Detection Based on Deep Learning Single-Stage Algorithm
-
摘要: 䗴化石是石炭纪、二叠纪重要的标准化石，其详细的鉴定工作对确定地质时代和划分石炭系‒二叠系具有重要意义.鉴于目前䗴类化石检测方法的局限性，提出一种基于深度学习单阶段算法的䗴类化石检测.以䗴类化石为研究对象，对原始模型进行分析，之后联合优化权重损失函数和BN层尺度因子的L1正则化等方式进行通道剪枝，再使用知识蒸馏使剪枝后模型恢复检测性能.实验结果表明，该方法可实现薄片图像中䗴类所在区域的定位和分类，平均精度均值达到98.1%，满足实时检测模型的要求，并且剪枝后参数量压缩了74.1%，解决了真实场景中存在的算力缺乏等问题.该方法能够有效保证䗴类化石的检测效果，同时扩展了该模型在嵌入式设备的适用范围，为深度学习在古生物化石图像的智能识别方面提供更多可能性.Abstract: Fusulinids are important standard fossils of the Carboniferous and Permian periods. The identification of fusulinids is significant for determining the geological age and delineating the Carboniferous-Permian stratigraphy. Considering the limitations of current fossil detection methods, a fusulinid detection method based on a deep learning single-stage algorithm is proposed. Taking fusulinids as the research object, the original model is improved by channel pruning by jointly optimizing the weight loss function and L1 regularization of the BN layer scale factor to compress the model size. Furthermore, the knowledge distillation is utilized to restore the detection performance of the pruned model. The experimental results show that the method can achieve the classification and localization of the fusulinids in the thin section images. The average accuracy reaches 98.1%, which meets the requirements of the real-time detection model. In addition, the number of model parameters is reduced by 74.1%, which solves the problems such as the lack of arithmetic power existing in real scenes. The method can effectively achieve the detection of fusulinids, while extending the applicability of the model to embedded devices and providing more possibilities for deep learning to perform intelligent recognition in paleontological fossil images.
-
Key words:
- fusulinid /
- deep learning /
- object detection /
- CarboniferousPermian /
- knowledge distillation /
- stratigraphy
-
图 1 中国华南和华北地区二叠系划分与对比(修改自沈树忠等，2019)
Fig. 1. Division and correlation of Permian in South and North China(modified after Shen et al., 2019)
图 2 基于深度学习单阶段算法的䗴类化石检测
Fig. 2. Fusulinid detection based on deep learning single-stage algorithm
图 5 稀疏训练前后$ \gamma $值分布变化
Fig. 5. Changes in the distribution of $ \gamma $ values before and after sparse training
图 6 知识蒸馏过程
Softmax（T=t/T=1）表示Softmax函数中温度系数为t或1；Soft Targets为softmax层的输出值；Soft predictions /Hard predictions表示经过Softmax（T=t）的输出值/T=1时Softmax的输出值；Loss Fn为损失函数
Fig. 6. Knowledge distillation process
图 7 䗴类化石显微图像部分样本
Fig. 7. Microscopic images of some fusulinids
a. Fusulina; b. Misellina; c. Neoschwagerina; d. Schwagerina; e. Triticites
图 10 不同微调策略下mAP曲线比较
Fig. 10. Comparison of mAP curves under different fine-tuning strategies
图 11 算法改进前后对化石显微图像的检测结果
a列. Fusulina；b列. Misellina；c列. Neoschwagerina；d列. Schwagerina；e. Triticites
Fig. 11. Detection results of microscopic images of fusulinids before and after algorithm improvement
表 1 本研究选取的䗴类化石
Table 1. The fusulinids selected for this study
属名 中文名 形状特征 含义 Fusulina 纺锤䗴属 壳纺锤形、粗纺锤形到圆柱形，旋壁多为四层式构成. 晚石炭世早期出现，代表地层为华北的本溪组. Triticites 麦粒䗴属 壳纺锤形到长纺锤形，具有旋脊，隔壁开始褶皱. 晚石炭世晚期出现，代表地层为华北太原组. Schwagerina 希瓦格䗴属 壳纺锤形、圆柱形和近球形，旋壁由致密层和蜂巢层构成. 晚石炭世至中二叠世末期，代表地层为华南的船山组. Neoschwagerina 新希瓦格䗴属 壳圆筒形到近球形，发育拟旋脊，具有副隔壁. 中二叠世，代表地层为华南的茅口组. Misellina 米斯䗴属 壳呈近球状，发育拟旋脊. 代表时期为早二叠世，代表地层为华南的栖霞组. 
下载: 导出CSV
表 2 实验参数
Table 2. Experimental parameters
参数名称 参数值 Image Size 640×640 Batch Size 64 Epochs 300 Learning_rate 0.001 NMS 0.5
下载: 导出CSV
表 3 主流目标检测模型性能对比
Table 3. Comparison of mainstream object detection model performance
模型 mAP(%) 计算量(GFLOPs) 模型大小(MB) 推理速度(ms/帧) SSD 79.80 87.45 92.64 34.00 Faster-RCNN 90.04 181.30 108.20 139.00 YOLOv3 95.9 155.0 234.0 39.5 YOLOv4 94.4 106.1 244.0 34.3 YOLOv5s 98.89 15.60 27.27 31.20 Ours 98.11 8.60 6.24 29.00
下载: 导出CSV
表 4 消融实验结果对比
Table 4. Comparison results of ablation experiments
模型 mAP(%) 通道数 参数量(M) 计算量GFLOPs 模型大小(MB) Baseline 98.89 9 632 7.71 15.6 27.27 Baseline+PR: 65% 75.3 4 308 1.57 8.6 6.24 Baseline+PR: 65%+KD 98.11 4 308 1.57 8.6 6.24
下载: 导出CSV
表 5 䗴类化石计数结果
Table 5. Results of fusulinids counts
检测类别 人工计数个数 䗴类化石计数个数 精确度（%） 䗴类化石 180 169 93.8
下载: 导出CSV
-
Bochkovskiy, A., Wang, C. Y., Liao, H. Y. M., 2020. YOLOv4: Optimal Speed and Accuracy of Object Detection. ArXiv, 2004.10934. https://arxiv.org/abs/2004.10934 Bu, J. J., He, W. H., Zhang, K. X., et al., 2020. Evolution of the Paleo-Asian Ocean: Evidences from Paleontology and Stratigraphy. Earth Science, 45(3): 711-727 (in Chinese with English abstract). Denil, M., Shakibi, B., Dinh, L., et al., 2013. Predicting Parameters in Deep Learning. ArXiv, 1306.0543. https://doi.org/10.48550/arXiv.1306.0543 Du, Y. S., Tong, J. N., 2009. Introduction to Palaeontology and Historical Geology. China University of Geosciences Press, Wuhan (in Chinese). Girshick, R., 2015. Fast R-CNN. 2015 IEEE International Conference on Computer Vision (ICCV). Santiago. https://doi.org/10.1109/ICCV.2015.169 Girshick, R., Donahue, J., Darrell, T., et al., 2014. Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation. 2014 IEEE Conference on Computer Vision and Pattern Recognition, Columbus. https://doi.org/10.1109/CVPR.2014.81 Guo, W., Lin, Y. T., Liu, G. H., 2003. Early Permian Rugose Coral Assemblage and Its Geological Significances in Xiwuqi of Inner Mongolia. Journal of Jilin University (Earth Science Edition), 33(4): 399-405 (in Chinese with English abstract). He, K. M., Zhang, X. Y., Ren, S. Q., et al., 2016. Deep Residual Learning for Image Recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas. https://doi.org/10.1109/CVPR.2016.90 Hinton, G., Vinyals, O., Dean, J., 2015. Distilling the Knowledge in a Neural Network. ArXiv, 1503.02531. https://doi.org/10.48550/arXiv.1503.02531 Hou, X. H., Feng, L., Zheng, M. P., et al., 2022. Recognition Method of Potassium-Rich Lithium Brine Reservoir in Nanyishan. Earth Science, 47(1): 45-55 (in Chinese with English abstract). Huang, Z., Bai, Z. Q., Chai, H., 2009. Identification of the 'Ambiguity Features' of the Conodont by the Artificial Neural Network. Geological Science and Technology Information, 28(3): 94-98 (in Chinese with English abstract). Li, P. F., Wang, H., Wu, Y. C., et al., 2022. Recognition of Dangerous Rock Mass and Seismic Risk Analysis of Highway Based on UAV Images. China Earthquake Engineering Journal, 44(4): 777-785 (in Chinese with English abstract). Liu, W., Anguelov, D., Erhan, D., et al., 2016. Ssd: Single Shot MultiBox Detector. 2016 European Conference on Computer Vision, Amsterdam. https://doi.org/10.1007/978-3-319-46448-0_2 Liu, X. Y., 2018. The Application of Image Recognition Technology in Paleontological Fossil Image (Dissertation). Jilin University, Changchun (in Chinese with English abstract). Liu, Z., Li, J. G., Shen, Z. Q., et al., 2017. Learning Efficient Convolutional Networks through Network Slimming. 2017 IEEE International Conference on Computer Vision (ICCV). Venice. https://doi.org/10.1109/ICCV.2017.298 Lowe, D. G., 2004. Distinctive Image Features from Scale-Invariant Keypoints. International Journal of Computer Vision, 60(2): 91-110. https://doi.org/10.1023/B:VISI.0000029664.99615.94 Mehta, R., Ozturk, C., 2018. Object Detection at 200 Frames Per Second. ArXiv, 1805.06361. https://doi.org/10.48550/arXiv.1805.06361 Peng, X. D., Zhang, M. S., Li, X. M., 1999. The Evolution of Paleozoic Tectonic Paleogeography in Jilin and Heilongjiang Orogenic Belt. World Geology, 18(3): 24-28 (in Chinese). Redmon, J., Divvala, S., Girshick, R., et al., 2016. You only Look Once: Unified, Real-Time Object Detection. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas. https://doi.org/10.1109/CVPR.2016.91 Redmon, J., Farhadi, A., 2018. Yolov3: An Incremental Improvement. arXiv, 1804.02767. https://doi.org/10.48550/arXiv.1804.02767 Ren, S. Q., He, K. M., Girshick, R., et al., 2015. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. arXiv, 1506.01497. https://doi.org/10.48550/arXiv.1506.01497 Shen, S. Z., Zhang, H., Zhang, Y. C., et al., 2019. Permian Integrative Stratigraphy and Timescale of China. Science in China (Series D), 49(1): 160-193 (in Chinese). Simonyan, K., Zisserman, A., 2014. Very Deep Convolutional Networks for Large-Scale Visual Recognition. ArXiv, 1409.1556. https://doi.org/10.48550/arXiv.1409.1556 Szegedy, C., Liu, W., Jia, Y. Q., et al., 2015. Going Deeper with Convolutions. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)., Boston. https://doi.org/10.1109/CVPR.2015.7298594 Weller, A. F., Harris, A. J., Ware, J. A., 2006. Artificial Neural Networks as Potential Classification Tools for Dinoflagellate Cyst Images: A Case Using the Self- Organizing Map Clustering Algorithm. Review of Palaeobotany and Palynology, 141(3-4): 287-302. https://doi.org/10.1016/j.revpalbo.2006.06.001 Yu, X. L., Ye, K., Du, C. J., et al., 2021. Microscopic Recognition of Micro Fossils in Carbonate Rocks Based on Convolutional Neural Network. Petroleum Geology & Experiment, 43(5): 880-885, 895 (in Chinese with English abstract). Yue, X., Hu, H., Jia, J. Z., 2019. A Method for Identifying Foraminifera Fossils Based on Deep Learning. Computer Knowledge and Technology, 15(27): 173, 178 (in Chinese). Zhang, Z. C., 1987. Correlation of the Shanxi Stage in North China and Related Fusulinid-Bearing Strata in Certain Areas. Regional Geology of China, 6(4): 359-363 (in Chinese). Zuo, R. G., Peng, Y., Li, T., et al., 2021. Challenges of Geological Prospecting Big Data Mining and Integration Using Deep Learning Algorithms. Earth Science, 46(1): 350-358 (in Chinese with English abstract). 卜建军, 何卫红, 张克信, 等, 2020. 古亚洲洋的演化: 来自古生物地层学方面的证据. 地球科学, 45(3): 711-727. doi: 10.3799/dqkx.2019.068 杜远生, 童金南, 2009. 古生物地史学概论. 武汉: 中国地质大学出版社. 郭伟, 林英铴, 刘广虎, 2003. 内蒙古西乌旗地区早二叠世皱纹珊瑚化石组合及其地质意义. 吉林大学学报(地球科学版), 33(4): 399-405. https://www.cnki.com.cn/Article/CJFDTOTAL-CCDZ200304002.htm 侯献华, 冯磊, 郑绵平, 等, 2022. 南翼山富钾锂卤水储层识别方法. 地球科学, 47(1): 45-55. doi: 10.3799/dqkx.2021.123 黄铮, 白志强, 柴华, 2009. 利用人工神经网络方法对牙形石鉴定中若干"模糊特征" 的数字化研究. 地质科技情报, 28(3): 94-98. https://www.cnki.com.cn/Article/CJFDTOTAL-DZKQ200903017.htm 李培锋, 王晖, 吴雨辰, 等, 2022. 基于无人机影像的危岩体识别及公路地震风险研究. 地震工程学报, 44(4): 777-785. https://www.cnki.com.cn/Article/CJFDTOTAL-ZBDZ202204004.htm 刘曦阳, 2018. 图像识别技术在古生物化石图像上的应用(硕士学位论文). 长春: 吉林大学. 彭向东, 张梅生, 李晓敏, 1999. 吉黑造山带古生代构造古地理演化. 世界地质, 18(3): 24-28. https://www.cnki.com.cn/Article/CJFDTOTAL-SJDZ199903004.htm 沈树忠, 张华, 张以春, 等, 2019. 中国二叠纪综合地层和时间框架. 中国科学(D辑), 49(1): 160-193. https://www.cnki.com.cn/Article/CJFDTOTAL-JDXK201901009.htm 余晓露, 叶恺, 杜崇娇, 等, 2021. 基于卷积神经网络的碳酸盐岩生物化石显微图像识别. 石油实验地质, 43(5): 880-885, 895. https://www.cnki.com.cn/Article/CJFDTOTAL-SYSD202105018.htm 岳翔, 呼和, 贾建忠, 2019. 一种基于深度学习的有孔虫化石识别方法. 电脑知识与技术, 15(27): 173, 178. https://www.cnki.com.cn/Article/CJFDTOTAL-DNZS201927076.htm 张志存, 1987. 华北山西阶与有关含(竹蜓)地层之对比. 中国区域地质, 6(4): 359-363. https://www.cnki.com.cn/Article/CJFDTOTAL-ZQYD198704008.htm 左仁广, 彭勇, 李童, 等, 2021. 基于深度学习的地质找矿大数据挖掘与集成的挑战. 地球科学, 46(1): 350-358. doi: 10.3799/dqkx.2020.111 -
点击查看大图
计量
- 文章访问数: 83
- HTML全文浏览量: 41
- PDF下载量: 10
- 被引次数: 0
