共享型轻量级卷积神经网络实现车辆外观识别 
下载： 1047次
Vehicle Appearance Recognition Using Shared Lightweight Convolutional Neural Networks
河北工业大学电子信息工程学院, 天津 300401
图 & 表
图 1. Fire模块结构
Fig. 1. Fire module structure
下载图片 查看原文
图 2. 结构对比。(a) FC;(b) GAP
Fig. 2. Structure comparison. (a) FC; (b) GAP
下载图片 查看原文
图 3. 网络结构。(a)无共享型;(b)部分共享型;(c)完全共享型
Fig. 3. Structure of networks. (a) No-shared; (b) partly-shared; (c) fully-shared
下载图片 查看原文
图 4. Opendata_VRID数据集中车辆图像样例
Fig. 4. Example of vehicle images in Opendata_VRID dataset
下载图片 查看原文
图 5. 数据集分布情况。(a)车辆型号;(b)车辆颜色
Fig. 5. Dataset distribution. (a) Vehicle type; (b) vehicle color
下载图片 查看原文
图 6. 不同“瘦身”程度SqueezeNet的结果对比。(a)颜色分类的训练损失;(b)颜色分类的验证准确率;(c)车型分类的训练损失;(d)车型分类的验证准确率
Fig. 6. Result comparison between different “slimming” SqueezeNet. (a) Training loss of color recognition; (b) validation accuracy of color recognition; (c) training loss of vehicle type recognition; (d) validation accuracy of vehicle type recognition
下载图片 查看原文
图 7. 车型识别结果对比。(a)训练损失;(b)验证准确率
Fig. 7. Result comparison for vehicle type recognition. (a) Training loss; (b) validation accuracy
下载图片 查看原文
图 8. 车辆颜色识别结果对比。(a)训练损失;(b)验证准确率
Fig. 8. Result comparison for vehicle color recognition. (a) Training loss; (b) validation accuracy
下载图片 查看原文
表 1三种不同“瘦身”程度的SqueezeNet结构
Table1. Three “slimming” SqueezeNet structures
| Number | Layer | Filter shape (N: number of categories) |
|---|
| SqueezeNet | Squeeze1 | Squeeze2 | Squeeze3 |
|---|
| 0 | Conv1 | 96 × 3 × 7 ×7 | 48 × 3 × 7 ×7 | 24 × 3 × 7 ×7 | 12 × 3 × 7 ×7 | | 1 | Fire2/Squeeze1×1 | 16 × 96 × 1 × 1 | 8 × 48 × 1 × 1 | 4 × 24 × 1 × 1 | 2 × 12 × 1 × 1 | | 2 | Fire2/Expand3×3 | 64 × 16 × 3 × 3 | 32 × 8 × 3 × 3 | 16 × 4 × 3 × 3 | 8 × 2 × 3 × 3 | | 3 | Fire3/Squeeze1×1 | 16 × 128 × 1 × 1 | 8 × 64 × 1 × 1 | 4 × 32 × 1 × 1 | 2 × 16 × 1 × 1 | | 4 | Fire3/Expand3×3 | 64 × 16 × 3 × 3 | 32 × 8 × 3 × 3 | 16 × 4 × 3 × 3 | 8 × 2 × 3 × 3 | | 5 | Fire4/Squeeze1×1 | 32 × 128 × 1 × 1 | 16 × 64 × 1 × 1 | 8 × 32 × 1 × 1 | 4 × 16 × 1 × 1 | | 6 | Fire4/Expand3×3 | 128 × 32 × 3 × 3 | 64 × 16 × 3 × 3 | 32 × 8 × 3 × 3 | 16 × 4 × 3 × 3 | | 7 | Fire5/Squeeze1×1 | 32 × 256 × 1 × 1 | 16 × 128 × 1 × 1 | 8 × 64 × 1 × 1 | 4 × 32 × 1 × 1 | | 8 | Fire5/Expand3×3 | 128 × 32 × 3 × 3 | 64 × 16 × 3 × 3 | 32 × 8 × 3 × 3 | 16 × 4 × 3 × 3 | | 9 | Fire6/Squeeze1×1 | 48 × 256 × 1 × 1 | 24 × 128 × 1 × 1 | 12 × 64 × 1 × 1 | 6 × 32 × 1 × 1 | | 10 | Fire6/Expand3×3 | 192 × 48 × 3 × 3 | 96 × 24 × 3 × 3 | 48 × 12 × 3 × 3 | 24 × 6 × 3 × 3 | | 11 | Fire7/Squeeze1×1 | 48 × 384 × 1 × 1 | 24 × 192 × 1 × 1 | 12 × 96 × 1 × 1 | 6 × 48 × 1 × 1 | | Number | Layer | Filter shape (N: number of categories) | | SqueezeNet | Squeeze1 | Squeeze2 | Squeeze3 | | 12 | Fire7/Expand3×3 | 192 × 48 × 3 × 3 | 96 × 24 × 3 × 3 | 48 × 12 × 3 × 3 | 24 × 6 × 3 × 3 | | 13 | Fire8/Squeeze1×1 | 64 × 384 × 1 × 1 | 32 × 192 × 1 × 1 | 16 × 96 × 1 × 1 | 8 × 48 × 1 × 1 | | 14 | Fire8/Expand3×3 | 256 × 64 × 3 × 3 | 128 × 32 × 3 × 3 | 64 × 16 × 3 × 3 | 32 × 8 × 3 × 3 | | 15 | Fire9/Squeeze1×1 | 64 × 512 × 1 × 1 | 32 × 256 × 1 × 1 | 16 × 128 × 1 × 1 | 8 × 64 × 1 × 1 | | 16 | Fire9/Expand3×3 | 256 × 64 × 3 × 3 | 128 × 32 × 3 × 3 | 64 × 16 × 3 × 3 | 32 × 8 × 3 × 3 | | 17 | Conv10 | N × 512 × 1 × 1 | N × 256 × 1 × 1 | N × 128 × 1 × 1 | N × 64 × 1 × 1 | | 18 | Global average pooling+SoftMax |
|
查看原文
表 29种网络结构的参数量对比
Table2. Comparison of number of parameters in nine network structures
| Network(basic network+sharing method) | Number of trainable parameters | Number of non-trainable parameters | Number of total parameters |
|---|
| Squeeze1+no-shared | 338825 | 838 | 339663 | | Squeeze1+partly-shared | 313993 | 678 | 314671 | | Squeeze1+fully-shared | 171873 | 438 | 172311 | | Squeeze2+no-shared | 90401 | 438 | 90839 | | Squeeze2+partly-shared | 82305 | 358 | 82663 | | Squeeze2+fully-shared | 46445 | 238 | 46683 | | Squeeze3+no-shared | 25469 | 238 | 25707 | | Squeeze3+partly-shared | 22501 | 198 | 22699 | | Squeeze3+fully-shared | 13371 | 138 | 13509 |
|
查看原文
表 3不同网络的结果对比
Table3. Result comparison between different networks
| Network(basic network+sharing method) | Testing accuracy/% | Time cost/ms |
|---|
| Color | Type |
|---|
| Squeeze1+no-shared | 98.6 | 99.2 | 6.89 | | Squeeze1+partly-shared | 98.5 | 99.2 | 4.99 | | Squeeze1+fully-shared | 98.5 | 99.1 | 4.42 | | Squeeze2+no-shared | 98.6 | 99.2 | 4.11 | | Squeeze2+partly-shared | 98.5 | 99.0 | 2.79 | | Squeeze2+fully-shared | 98.4 | 98.8 | 2.48 | | Squeeze3+no-shared | 98.3 | 99.0 | 3.06 | | Squeeze3+partly-shared | 98.0 | 98.3 | 2.01 | | Squeeze3+fully-shared | 97.5 | 97.0 | 1.68 |
|
查看原文
康晴, 赵红东, 杨东旭. 共享型轻量级卷积神经网络实现车辆外观识别[J]. 激光与光电子学进展, 2021, 58(2): 0210013. Qing Kang, Hongdong Zhao, Dongxu Yang. Vehicle Appearance Recognition Using Shared Lightweight Convolutional Neural Networks[J]. Laser & Optoelectronics Progress, 2021, 58(2): 0210013.
PDF全文
