基于二值语义分割网络的遥感建筑物检测 
下载： 989次
Remote Sensing Building Detection Based on Binarized Semantic Segmentation
1 中国科学院红外探测与成像技术重点实验室, 上海 200083
2 中国科学院上海技术物理研究所, 上海 200083
3 中国科学院大学, 北京 100049
图 & 表
图 1. 传统的神经网络训练流程图
Fig. 1. Flow chart for training of traditional neural network
下载图片 查看原文
图 2. 二值化神经网络的训练流程图
Fig. 2. Flow chart for training of binarized neural network
下载图片 查看原文
图 3. 三种网络对比示意图
Fig. 3. Comparison of three neural networks
下载图片 查看原文
图 4. 三种网络结构和详细参数。(a) FU-Net;(b) GBU-Net;(c) MBU-Net
Fig. 4. Details and architectures of three networks. (a) FU-Net; (b) GBU-Net; (c) MBU-Net
下载图片 查看原文
图 5. 遥感图片上的建筑物及其对应标注。(a) 卫星遥感图片;(b) 标注图片
Fig. 5. Buildings in satellite remote sensing image and their corresponding labels. (a) Satellite remote sensing image; (b) labeled image
下载图片 查看原文
图 6. 部分训练数据图片
Fig. 6. Part samples of satellite remote sensing data for training
下载图片 查看原文
图 7. MBU-Net卷积核组参数可视化结果
Fig. 7. Visualization result of convolution kernel group parameters of MBU-Net
下载图片 查看原文
图 8. MBU-Net中间层可视化输入图
Fig. 8. Input image for visualizing of MBU-Net interlayer
下载图片 查看原文
图 9. 第三层30张特征图的可视化结果
Fig. 9. Visualization results of 30 feature maps in third layer
下载图片 查看原文
图 10. 五种网络的训练损失变化曲线
Fig. 10. Trends of loss of five networks
下载图片 查看原文
图 11. 五种网络的训练准确率变化曲线
Fig. 11. Trends of accuracy of five networks
下载图片 查看原文
图 12. 测试集中的部分输入图片及其对应的输出预测
Fig. 12. Partial input images and output prediction on test set
下载图片 查看原文
表 1MBU-Net网络层详细的输入输出参数表
Table1. Input and output parameters of MBU-Net layer
| Convolution type | Input | Kernel size | Stride | Padding | Output |
|---|
| BinaryConv2D | 256×256×3 | 3×3×64 | 1 | Same | 256×256×64 | | BinaryConv2D | 256×256×64 | 3×3×64 | 1 | Same | 256×256×64 | | Pool | 256×256×64 | 2×2 | 1 | | 128×128×64 | | BinaryConv2D | 128×128×64 | 3×3×128 | 1 | Same | 128×128×128 | | BinaryConv2D | 128×128×128 | 3×3×128 | 1 | Same | 128×128×128 | | Pool | 128×128×128 | 2×2 | 1 | | 64×64×128 | | BinaryConv2D | 64×64×128 | 3×3×256 | 1 | Same | 64×64×256 | | BinaryConv2D | 64×64×256 | 3×3×256 | 1 | Same | 64×64×256 | | Pool | 64×64×256 | 2×2 | | | 32×32×256 | | BinaryConv2D | 32×32×256 | 3×3×512 | 1 | Same | 32×32×512 | | BinaryConv2D | 32×32×512 | 3×3×512 | 1 | Same | 32×32×512 | | Pool | 32×32×512 | 2×2 | | | 16×16×512 | | ︙ | ︙ | ︙ | ︙ | ︙ | ︙ | | BinaryConv2D | 256×256×128 | 3×3×64 | 1 | Same | 256×256×64 | | BinaryConv2D | 256×256×64 | 3×3×16 | 1 | Same | 256×256×16 | | Conv2D | 256×256×16 | 1×1×1 | 1 | Same | 256×256×1 |
|
查看原文
表 2五种网络的训练损失值和准确率
Table2. Results of training loss and accuracy of five networks
| Algorithm | Training loss | Training accuracy /% |
|---|
| FU-Net | 0.01 | 99.56 | | GBU-Net | 0.41 | 87.73 | | MBU-Net | 0.16 | 93.94 | | Deeplab | 0.01 | 99.70 | | ENet | 0.23 | 92.21 |
|
查看原文
表 3五种网络在测试集上的各项测试指标
Table3. Test indexes of five networks on test set
| Algorithm | /% | P /% | R /% | /% | Memory /MB | Time /ms |
|---|
| FU-Net | 81.88 | 77.07 | 66.20 | 71.22 | 96.60 | 357 | | GBU-Net | 74.74 | 84.41 | 54.22 | 66.03 | 3.02 | 47 | | MBU-Net | 82.33 | 82.76 | 65.54 | 73.15 | 3.02 | 49 | | Deeplab | 82.03 | 78.63 | 67.11 | 72.41 | 123.47 | 489 | | ENet | 78.49 | 85.37 | 56.19 | 67.77 | 13.73 | 107 |
|
查看原文
表 4不同类型参数加法和乘法运算操作的能耗对比[32]
Table4. Comparison of energy consumption of addition and multiplication operations for different types of parameters[32]
| Operation | Energy consumption /pJ |
|---|
| Multiplication | Addition |
|---|
| 8 bit integer | 0.20 | 0.03 | | 32 bit integer | 3.10 | 0.10 | | 16 bit floating point | 1.10 | 0.40 | | 32 bit floating point | 3.70 | 0.90 |
|
查看原文
朱天佑, 董峰, 龚惠兴. 基于二值语义分割网络的遥感建筑物检测[J]. 光学学报, 2019, 39(12): 1228002. Tianyou Zhu, Feng Dong, Huixing Gong. Remote Sensing Building Detection Based on Binarized Semantic Segmentation[J]. Acta Optica Sinica, 2019, 39(12): 1228002.
PDF全文
