基于卷积神经网络的混合颗粒分类法研究 下载: 841次
Method for Mixed-Particle Classification Based on Convolutional Neural Network
上海理工大学能源与动力工程学院, 上海 200093
图 & 表
图 1. 混合颗粒分类的CNN结构
Fig. 1. CNN structure of mixed-particle classification
下载图片 查看原文
图 2. RPN网络结构。(a)锚框尺度设置;(b)锚框比例设置
Fig. 2. RPN network structure. (a) Anchor frame scale setting; (b) anchor frame ratio setting
下载图片 查看原文
图 3. 混合颗粒的分类算法流程
Fig. 3. Flow chart of mixed-particle classification algorithm
下载图片 查看原文
图 4. 不同类型的颗粒。(a)球形颗粒;(b)长条形颗粒;(c)非规则颗粒
Fig. 4. Examples of different types of particles. (a) Spherical particles; (b) elongated particles; (c) irregular particles
下载图片 查看原文
图 5. 混合颗粒测量系统
Fig. 5. Measuring system of mixed particles
下载图片 查看原文
图 6. 图像处理流程
Fig. 6. Flow chart of image processing
下载图片 查看原文
图 7. 不同方法处理后混合颗粒的图像。(a)混合颗粒;(b)维纳滤波;(c)二值化和孔洞填充;(d) watershed分割;(e)手动精分割
Fig. 7. Images of mixed particles processed by different methods. (a) Mixed particles; (b) Wiener filtering; (c) binarization and hole filling; (d) watershed segmentation; (e) manually fine segmentation
下载图片 查看原文
图 8. 不同分类方法得到的颗粒计数结果
Fig. 8. Particle counting results obtained by different classification methods
下载图片 查看原文
图 9. 不同分类方法的检测准确率
Fig. 9. Detection accuracy of different classification methods
下载图片 查看原文
图 10. 不同分类方法得到的颗粒的当量直径累积分布
Fig. 10. Cumulative distributions of equivalent diameters of particles obtained by different classification methods
下载图片 查看原文
图 11. 不同分类方法得到的长条形颗粒的长宽比累积分布
Fig. 11. Cumulative distributions of aspect ratios for elongated particles obtained by different classification methods
下载图片 查看原文
图 12. 不同分类方法得到的球形颗粒的长宽比累积分布
Fig. 12. Cumulative distributions of aspect ratios for spherical particles obtained by different classification methods
下载图片 查看原文
图 13. 不同分类方法得到的非规则颗粒的长宽比累积分布
Fig. 13. Cumulative distributions of aspect ratios for irregular particles obtained by different classification methods
下载图片 查看原文
表 1颗粒的特征描述子
Table1. Feature descriptors of particles
Parameter | Symbol | Description | Category |
---|
Perimeter | P | The distance around the boundary of the region | | Area | A | The actual number of pixels in the region | | Equivalent diameter | Deq | Deq=2 | 1 | Major axis | L | The major axis of the external ellipse | | Minor axis | S | The minor axis of the external ellipse | | Circularity | C | C= | | Aspect ratio | AR | The aspect ratio of minimum bounding rectangle | | Boundary irregularity | Birr | Birr=2π | 2 | Uniformity | U | The ratio of the outer rectangle to the outer convex polygon | | Angular point | AP | The number of concave and convex points | |
|
查看原文
表 2conv3_x和conv4_x的结构配置
Table2. Structure configurations of conv3_x and conv4_x
Stage | Output size /(pixel×pixel) | Block structure | Block count |
---|
conv3_x | 28×28 | | 3 | conv4_x | 14×14 | | 20 |
|
查看原文
表 3不同分类方法得到的颗粒的测量尺寸
Table3. Particle sizes measured by different classification methods
Method | Diameter of spherical particles /μm | Diameter of irregular particles /μm | Diameter of elongated particles /μm | | |
---|
| | | | | | | | |
---|
Ground truth | 111.9 | 122.5 | 131.1 | 98.5 | 136.4 | 172.9 | 74.4 | 96.0 | 137.6 | CNN | 110.0 | 118.5 | 126.8 | 103.3 | 136.4 | 171.2 | 74.3 | 95.8 | 135.9 | SVM_1 | 110.0 | 117.2 | 129.7 | 84.6 | 120.9 | 160.4 | 73.3 | 94.4 | 138.4 | SVM_2 | 110.5 | 117.3 | 131.3 | 75.7 | 117.8 | 159.4 | 69.7 | 93.2 | 139.0 | BP | 105.3 | 116.1 | 137.3 | 84.2 | 112.0 | 157.7 | 76.0 | 101.1 | 147.4 |
|
查看原文
蔡杨, 苏明旭, 蔡小舒. 基于卷积神经网络的混合颗粒分类法研究[J]. 光学学报, 2019, 39(7): 0712002. Yang Cai, Mingxu Su, Xiaoshu Cai. Method for Mixed-Particle Classification Based on Convolutional Neural Network[J]. Acta Optica Sinica, 2019, 39(7): 0712002.