Laser & Optoelectronics Progress, Volume. 57, Issue 20, 201014(2020)
Intelligent Domestic Garbage Recognition Based on Faster RCNN
In this paper, we presented the Intelligent Domestic Garbage Recognition using Faster RCNN to realize high-precision identification of domestic garbage. Specifically, 6 kinds of domestic garbage were selected to build the dataset. The data augmentation technique was adopted to expand the quantity and category of the targets, and improve balance on the size of the targets. Moreover, we used three different types of backbone networks including VGG-16, Res101, and MobileNet_v1 to analyze and compare the accuracy, speed, and generalization performance. The research used end-to-end training network finely tuned by the special layer, and carried out enhanced training on low recognition rate samples to obtain a minimum mean average precision (mAP) of 92.85%. Subsequently, we captured three typical errors and optimized from the misidentified samples, and thus the highest recognition mAP increased to 99.23%. To analyze the generalization performance of different backbone networks embedded in the algorithm, we built a dataset with 816 pictures derivatized from the different backgrounds and used it to test the impact of changing the background on garbage detection. As a result, we found that the complex backgrounds from surrounding garbage put the greatest impact on detection accuracy. Thus, the generalization performance takes the same trend as convergence performance, which changes Res101, VGG-16, MobileNet_v1 from good to bad. Therefore, the setting method of the optimal probability threshold for algorithm detection was analyzed and obtained based on the principles of the high-precision requirement for recyclable garbage and high recall requirements for hazardous garbage.
Get Citation
Copy Citation Text
Canhua Wen, Jia Li, Xue Dong. Intelligent Domestic Garbage Recognition Based on Faster RCNN[J]. Laser & Optoelectronics Progress, 2020, 57(20): 201014
Category: Image Processing
Received: Jan. 13, 2020
Accepted: Mar. 9, 2020
Published Online: Oct. 13, 2020
The Author Email: Li Jia (canhuamail@163.com)