GGCN:基于GPU的高光谱图像分类算法 
下载： 1049次
GGCN: GPU-Based Hyperspectral Image Classification Algorithm
1 上海海洋大学信息学院, 201306
2 上海电力大学电子与信息工程学院, 上海 200090
图 & 表
图 1. Cube-CNN-SVM模型框架
Fig. 1. Cube-CNN-SVM model framework
下载图片 查看原文
图 2. 卷积运算的定义形式
Fig. 2. Definition form of convolution operation
下载图片 查看原文
图 3. 卷积运算的矩阵乘形式
Fig. 3. Matrix multiplication form of convolution operation
下载图片 查看原文
图 4. 图像预处理以及卷积运算
Fig. 4. Image preprocessing and convolution operation
下载图片 查看原文
图 5. 模型训练损失和精度变化曲线。(a)损失;(b)精度
Fig. 5. Model training loss and accuracy variation. (a) Loss; (b) accuracy
下载图片 查看原文
图 6. 不同数量卷积层加速比的变化
Fig. 6. Changes in the speedup ratio of different numbers of convolution layers
下载图片 查看原文
表 1算法伪代码
Table1. Algorithm pseudocode
| Algorithm:GGCN |
|---|
| Input: Hyperspectral image1, Data preprocessing: processing <<>>i-th iteration: Forward propagation2, Convolutional: convol <<< gridsize, blocksize, 0, stream>>>3, Pooling: maxpooling <<< gridsize, blocksize, 0, stream>>>4, Fully connected: fullyconnected <<< gridsize, blocksize, 0, stream>>>5, Output: output <<>>6, Copy classification results to CPU to calculate the loss: 7, Copy data: cudaMemcpy()8, Calculate the loss: lossfunction()Backward propagation9, Output: bp_output <<< gridsize, blocksize, 0, stream>>>10, Fully Connected: bp_fullyconnected<<< gridsize, blocksize,0,stream>>>11, Pooling: bp_maxpooling <<>>12, Convolutional: bp_update_kernel <<< gridsize, blocksize, 0, stream>>>OutputEnd |
|
查看原文
表 2遥感数据集信息
Table2. Information of the remote sensing datasets
| Dataset | Sensor | Class number | Dimension | Top 5 classes | Size /MB |
|---|
| KSC | AVIRIS | 13 | 512 × 614×176 | Water, scrub, spartna-marsh,mud-flats, salt-marsh | 56.8 | | PU | POSIS | 9 | 610×340×103 | Meadows, asphalt, bare-soil,self-blocking bricks, trees | 33.2 | | Indian Pines | AVIRIS | 16 | 145×145×224 | Soybean-mintill, corn-notill, woods,soybean-notill, corn-mintill | 5.7 |
|
查看原文
表 3不同数据预处理方式时间消耗对比
Table3. Comparison of time consumption of different data preprocessing methods
| Dataset | Neighbor pixel extract strategy | Time /s |
|---|
| CPU | PNPE | G-PNPE |
|---|
| KSC | 1P4N8N | 2.654.896.21 | 0.450.881.12 | 0.510.921.22 | | PU | 1P4N8N | 2.213.303.87 | 0.310.520.66 | 0.330.490.71 | | Indian Pines | 1P4N8N | 1.05 1.652.17 | 0.170.210.28 | 0.180.200.26 |
|
查看原文
表 4不同分类模型平均训练时间和加速比
Table4. Comparison of running time and speedup of different classification models
| Dataset | Method | Time /s | Speedup ratio |
|---|
| MBGD(batchsize is 10) | MBGD(batchsize is 100) |
|---|
| KSC | Cube-CNN-SVMGCNGGCN | 23123.623487.342834.01 | 23012.493322.342598.78 | 1.06.68.2 | | PU | Cube-CNN-SVMGCNGGCN | 2231.23351.46286.22 | 2187.75338.11230.06 | 1.06.37.8 | | Indian Pines | Cube-CNN-SVMGCNGGCN | 453.62107.3484.01 | 422.49102.3480.78 | 1.04.25.1 |
|
查看原文
表 5不同分类模型准确率
Table5. Accuracy of different classification models
| Dataset | Method | Accuracy /% |
|---|
| MBGD(batchsize is 10) | MBGD(batchsize is 100) |
|---|
| KSC | Cube-CNN-SVMGCNGGCN | 93.7893.3393.67 | 93.4793.1293.92 | | PU | Cube-CNN-VMGCNGGCN | 96.6796.2396.34 | 95.2195.6195.69 | | Indian Pines | Cube-CNN-SVMGCNGGCN | 94.7894.7394.87 | 94.6794.5294.42 |
|
查看原文
表 6改进模型与原模型各层级时间占比
Table6. Ratio of time between the improved model and the original model
| Layer | Percentage /% |
|---|
| GCN | GGCN |
|---|
| Preprocessing | 1.0 | 2.4 | | Convolution | 38.2 | 28.0 | | Pooling | 2.4 | 5.6 | | Fully connection | 27.1 | 24.0 | | Output | 19.0 | 22.0 | | Others | 13.3 | 18.0 |
|
查看原文
张明华, 邹亚晴, 宋巍, 黄冬梅, 刘智翔. GGCN:基于GPU的高光谱图像分类算法[J]. 激光与光电子学进展, 2020, 57(20): 201101. Minghua Zhang, Yaqing Zou, Wei Song, Dongmei Huang, Zhixiang Liu. GGCN: GPU-Based Hyperspectral Image Classification Algorithm[J]. Laser & Optoelectronics Progress, 2020, 57(20): 201101.
PDF全文
