基于卷积神经网络的低参数量实时图像分割算法 下载: 1195次
Low-Parameter Real-Time Image Segmentation Algorithm Based on Convolutional Neural Network
西南交通大学信息科学与技术学院, 四川 成都 611756
图 & 表
图 1. 卷积核。(a)经典卷积核;(b)空洞卷积核Rrate=2;(c)空洞卷积核Rrate=3
Fig. 1. Convolution kernel. (a) Classical convolution kernel; (b) dilated convolution kernel Rrate=2; (c) dilated convolution kernel Rrate=3
下载图片 查看原文
图 2. Atrous-Fire模块结构
Fig. 2. Atrous-Fire modular structure
下载图片 查看原文
图 3. 空洞卷积核与初始特征图。(a)锯齿结构卷积核;(b)无栅格特征图;(c)栅格特征图
Fig. 3. Dilated convolution kernel and initial characteristic graphs. (a) Sawtooth structure convolution kernel; (b) no grid feature graph; (c) grid feature graph
下载图片 查看原文
图 4. Atrous-squeezeseg网络结构
Fig. 4. Network structure of Atrous-squeezeseg
下载图片 查看原文
图 5. 训练损失值曲线图
Fig. 5. Training loss value curves
下载图片 查看原文
图 6. 验证损失值曲线
Fig. 6. Validation loss value curves
下载图片 查看原文
图 7. ADE20K效果对比图。(a)原图像;(b)分割标注图;(c)所提算法;(d) Squeezeseg+FCN; (e) VGG16+FCN;(f) SqueezeNet+FCN;(g)无空洞;(h)无批量标准化处理
Fig. 7. Effect comparison of ADE20K. (a) Original images; (b) ground truth; (c) proposed algorithm; (d) Squeezeseg+FCN; (e) VGG16+FCN; (f) SqueezeNet+FCN; (g) without dilated; (h) without BN
下载图片 查看原文
图 8. PASCAL VOC效果对比图。(a)原图像;(b)分割标注图;(c)所提算法;(d) Squeezeseg+FCN; (e) VGG16+FCN;(f) SqueezeNet+FCN;(g)无空洞;(h)无批量标准化处理
Fig. 8. Effect comparison of PASCAL VOC. (a) Original images; (b) ground truth; (c) proposed algorithm; (d) Squeezeseg+FCN; (e) VGG16+FCN; (f) SqueezeNet+FCN; (g) without dilated; (h) without BN
下载图片 查看原文
表 1编码器参数
Table1. Encoder parameters
Layer name | Output size | Squeeze(S1) | Expand(E1/E3) | Rrate |
---|
Input image | 224×224×3 | | | | Conv1 | 112×112×64 | | | | Maxpool1 | 56×56×64 | | | | Atrous-Fire1 (3×) | 56×56×256 | 16 | 32 | 2/5/7 | Maxpool2 | 28×28×256 | | | | Atrous-Fire2 (3×) | 28×28×256 | 32 | 64 | 2/3/5 | Maxpool3 | 14×14×256 | | | | Atrous-Fire3 (3×) | 14×14×256 | 64 | 128 | 2/3/5 | Atrous-Fire4 (2×) | 14×14×512 | 128 | 256 | 1/2 | Atrous-Fire4 (2×) | 14×14×512 | 128 | 256 | 1/1 |
|
查看原文
表 2不同语义分割模型的参数量与MIU
Table2. Number of parameters of different semantic segmentation models and MIU
Method | Number ofparameters | MIU | Building | Sky | Car | Tree | Road | Person | Floor | Wall |
---|
Atrous-squeezeseg | 21.09 | 62.9 | 67.5 | 84.0 | 61.4 | 58.1 | 64.7 | 49.1 | 60.4 | 58.5 | Squeezeseg+FCN | 54.65 | 55.9 | 61.8 | 85.8 | 48.4 | 51.8 | 61.5 | 32.3 | 53.8 | 52.2 | VGG16+FCN | 66.21 | 63.2 | 68.3 | 86.8 | 61.1 | 58.2 | 66.0 | 48.5 | 58.3 | 57.4 | SqueezeNet+FCN | 54.65 | 50.5 | 46.7 | 83.8 | 44.7 | 51.8 | 55.5 | 28.0 | 47.3 | 46.7 | Atrous-squeezeseg(without dilated) | 21.09 | 50.6 | 51.1 | 83.5 | 41.2 | 43.8 | 53.8 | 29.7 | 51.4 | 50.1 | Atrous-squeezeseg(without BN) | 21.09 | 51.6 | 51.3 | 83.5 | 43.6 | 45.8 | 58.0 | 29.8 | 50.1 | 51.3 |
|
查看原文
表 3PA与不同设备中模型的FPS值
Table3. PA and FPS of model in different devices
Method | FPS /(frame·s-1) | PA /% |
---|
| GTX 1080Ti | NVIDIA TX2 |
---|
Atrous-squeezeseg | 45.3 | 8.3 | 59.5 | Squeezeseg+FCN | 39.5 | 4.2 | 59.3 | VGG16+FCN | 29.6 | 1.9 | 59.8 | SqueezeNet+FCN | 46.6 | 4.5 | 55.6 | Atrous-squeezeseg(without dilated) | 45.6 | 8.4 | 56.1 | Atrous-squeezeseg(without BN) | 56.2 | 9.2 | 57.3 |
|
查看原文
谭光鸿, 侯进, 韩雁鹏, 罗朔. 基于卷积神经网络的低参数量实时图像分割算法[J]. 激光与光电子学进展, 2019, 56(9): 091003. Guanghong Tan, Jin Hou, Yanpeng Han, Shuo Luo. Low-Parameter Real-Time Image Segmentation Algorithm Based on Convolutional Neural Network[J]. Laser & Optoelectronics Progress, 2019, 56(9): 091003.