基于Faster RCNN的生活垃圾智能识别 
下载： 962次
Intelligent Domestic Garbage Recognition Based on Faster RCNN
上海交通大学中英国际低碳学院, 上海 201306
图 & 表
图 1. 实验装置
Fig. 1. Experimental equipment
下载图片 查看原文
图 2. 采集的各类别图片典型样本
Fig. 2. Typical image samples from each class
下载图片 查看原文
图 3. Faster RCNN算法结构
Fig. 3. Network structure of Faster RCNN
下载图片 查看原文
图 4. 结合困难样本增强及特异层微调的Faster RCNN算法训练步骤
Fig. 4. Faster RCNN train process combined with hard samples enhancement and special layer fine-tuning
下载图片 查看原文
图 5. 各网络总损失收敛情况及各训练步骤在测试集上的mAP
Fig. 5. Total loss convergence and mAP of test dataset during training procedure
下载图片 查看原文
图 9. MobileNet_v1概率阈值判定曲线
Fig. 9. Probability threshold decision curve on MobileNet_v1
下载图片 查看原文
表 1数据集各类别目标数量
Table1. Object quantity on garbage dataset
| Dataset | Metal | Plastic | Carton | Battery | Bulb | Pill | Total |
|---|
| Original dataset | 1321 | 1392 | 807 | 1058 | 1129 | 1402 | 7109 | | Augmented train dataset | 2429 | 2435 | 1938 | 2389 | 2398 | 2528 | 14117 | | Augmented test dataset | 630 | 585 | 483 | 597 | 617 | 630 | 3542 | | Augmented dataset | 3059 | 3020 | 2421 | 2986 | 3015 | 3158 | 17659 |
|
查看原文
表 2不同网络参数量、计算量及网络层数
Table2. Number of parameters, FLOPs and layers for different networks
| Network | Number ofparameters /107 | Number ofFLOPs /1010 | Layernumbers |
|---|
| VGG-16 | 136.79 | 166.37 | 20 | | Res101 | 47.21 | 167.25 | 105 | | MobileNet_v1 | 5.61 | 19.02 | 32 |
|
查看原文
表 3各网络训练集(TR)及测试集(TE)测试结果
Table3. Network results on train dataset (TR) and test dataset (TE)
| Backbonenetwork | AP | mAP | OptimizedmAP | Detection speed /(frame·s-1) |
|---|
| Metal | Plastic | Carton | Battery | Pill | Bulb |
|---|
| Res101 | TR | 1.0 | 0.9996 | 0.9985 | 0.9996 | 0.9973 | 1.0 | 0.9992 | 0.9993 | ~7 | | TE | 0.9770 | 0.9597 | 0.9817 | 0.9695 | 0.9728 | 0.9811 | 0.9736 | 0.9857 | | | VGG-16 | TR | 1.0 | 0.9997 | 0.9997 | 0.9996 | 0.9970 | 1.0 | 0.9993 | 0.9992 | ~9 | | TE | 0.9758 | 0.9639 | 0.9866 | 0.9835 | 0.9813 | 0.9853 | 0.9794 | 0.9923 | | | MobileNet_v1 | TR | 0.9817 | 0.9715 | 0.9732 | 0.9851 | 0.9831 | 0.9879 | 0.9804 | 0.9833 | ~20 | | TE | 0.9139 | 0.8737 | 0.9204 | 0.9408 | 0.9671 | 0.9554 | 0.9285 | 0.9490 | |
|
查看原文
表 4背景数据集测试结果
Table4. Test results on background dataset
| Backbone network | Original mAP | Status | mAP under different background types |
|---|
| Pure color | Texture | Garbage |
|---|
| Res101 | 1.0 | Before re-training | 0.9913 | 0.9222 | 0.9050 | | | After re-training | 1.0 | 1.0 | 1.0 | | VGG-16 | 1.0 | Before re-training | 0.9901 | 0.8835 | 0.6494 | | | After re-training | 1.0 | 1.0 | 1.0 | | MobileNet_v1 | 0.9917 | Before re-training | 0.9691 | 0.7433 | 0.4204 | | | After re-training | 0.9999 | 0.9933 | 0.9793 |
|
查看原文
表 5各网络最优阈值下测试集精度及召回率
Table5. Precision and recall on test dataset under optimal threshold of each network
| Backbone network(P1,P2) | Parameter | Recyclable garbage | Hazardous garbage |
|---|
| Metal | Plastic | Carton | Mean | Battery | Pill | Bulb | Mean |
|---|
| Res101(0.76, 0.24) | Precision | 0.9796 | 0.9662 | 0.9834 | 0.9764 | 0.9497 | 0.9792 | 0.9698 | 0.9662 | | Recall | 0.9889 | 0.9778 | 0.9834 | 0.9834 | 0.9799 | 0.9714 | 0.9887 | 0.9800 | | VGG-16(0.62, 0.38) | Precision | 0.9583 | 0.9487 | 0.9775 | 0.9615 | 0.9657 | 0.9658 | 0.9534 | 0.9491 | | Recall | 0.9857 | 0.9795 | 0.9876 | 0.9843 | 0.9899 | 0.9873 | 0.9951 | 0.9908 | | MobileNet_v1(0.56, 0.44) | Precision | 0.8943 | 0.9235 | 0.9109 | 0.9096 | 0.9266 | 0.9636 | 0.8867 | 0.9256 | | Recall | 0.9609 | 0.9322 | 0.9654 | 0.9528 | 0.9728 | 0.9739 | 0.9854 | 0.9774 |
|
查看原文
文灿华, 李佳, 董雪. 基于Faster RCNN的生活垃圾智能识别[J]. 激光与光电子学进展, 2020, 57(20): 201014. Canhua Wen, Jia Li, Xue Dong. Intelligent Domestic Garbage Recognition Based on Faster RCNN[J]. Laser & Optoelectronics Progress, 2020, 57(20): 201014.
PDF全文
