论文笔记[8] Seismic Waveform Classification and First-Break Picking Using CNN
2018-03-06 00:32
344 查看
论文笔记[8] Seismic Waveform Classification and First-Break Picking Using Convolution Neural Networks
这篇文章是用CNN做一个波形分类以及初至拾取。初至,first-break,简写为FB,在FB附近和non-FB附近的subimage的波形特征是不同的。初至拾取这个任务可以人工也可以automatic,现有的automatic method有STA/LTA ,autoregressive techniques,time-frequency transform,以及高阶统计量等,但是它们常常not adaptive,即适应性差,且对噪声敏感。另外,这些多单道处理,忽略了道间的空间特征。中间的部分讲CNN,基本都是教材内容,可能是地震行业中CNN的应用还不是很多。优化的loss函数就是l2范数,之后有个define的discriminant score function,如下:
由于这个把两类,即FB和non-FB分别用vector (1 0) 和 (0 1)表示,又是用的sigmoid函数,所以输出的o=(o1,o2)如果按照sigmoid和softmax的关系来说,只要o是归一化的,那么o1和o2相加就是1,所以这里的F作为一个判别式实际上就相当于两倍的o2的绝对值,文中用1作为F分类的边界,其实就是sigmoid函数在(0.5,0.5)处的分界,也就是相当于逻辑斯蒂回归中的中间点。所以其实可以直接两类的softmax取最大即可,F意义不大。
test或者说validation 的时候,对每个点取一个subimage,并且做二分类,这样所有的点都得到了一个binary的结果,是FB或者non-FB,然后用first local minimum,对每一道都做一遍,这样去掉一些false FB,然后,用median filter,考虑上空间信息。
下面是三个QC rules: 1) the separability of the referenced FB classification appearance represented by the carefully manually picked FB and the other classification appearances (false FB and non-FB classification); 2) the matching degree between the CNNs-based automatic picking FB and the manually picked FB; and 3) the quantity and randomness of false FBs are considered to qualitatively evaluate the quality of the chosen CNN input samples and output classification or the trained CNNs structure.
实验中只用了两层Conv,一个pooling,一个FC。在合成和实际数据上结果如下:
2018年03月06日00:31:54
相关文章推荐
- recording and play using Waveform audio interface
- 论文笔记:Research and Implementation of a Multi-label Learning Algorithm for Chinese Text Classification
- 论文笔记:Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks
- 论文笔记:A survey of recent advances in CNN-based single image crowd counting and density estimation
- 论文笔记:Histopathological Image Classification Using Discriminative Feature-Oriented Dictionary Learnin
- 论文笔记 A Large Contextual Dataset for Classification,Detection and Counting of Cars with Deep Learning
- Region-based Convolutional Networks for Accurate Object Detection and Segmentation----R-CNN论文笔记
- 论文阅读笔记:R-CNN:Rich feature hierarchies for accurate object detection and semantic segmentation
- [论文笔记] Post-copy based live virtual machine migration using adaptive pre-paging and dynamic self-ballooning (VEE, 2009)
- [论文笔记] On Construction of Cloud IaaS for VM Live Migration Using KVM and OpenNebula (ICA3PP, 2012)
- 笔记-论文-OverFeat Integrated Recognition, Localization and Detection using Convolutional Networks
- 论文读书笔记-Action unit classification using active appearance models and conditional random fields
- 论文笔记 | Exploit All the Layers: Fast and Accurate CNN Object Detector with SDP and CRC
- Appearance-and-Relation Networks for Video Classification论文笔记 (ARTnet)
- [深度学习论文笔记][Object Localization] OverFeat: Integrated Recognition, Localization and Detection using C
- [置顶] 论文笔记--Appearance-and-Relation Networks for Video Classification
- 论文笔记 | CNN-RNN:A Unified Framework for Multi-label Image Classification
- 论文笔记-MTCNN-Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Network
- 多尺度R-CNN论文笔记(3): HyperNet: Towards Accurate Region Proposal Generation and Joint Object Detection
- 【论文笔记】Relation Classification via Multi-Level Attention CNNs