Waste Sorting System Using Binarized Neural Network
Mingmei Wu, He Wei, Yingcheng Lin, Li Li, Songhong Liang, Xichuan Zhou · 2020
Garbage classification is of great importance to reduce environmental pollution and resource waste. For the sake of realizing efficient waste sorting and treatment, we propose a low latency automatic waste sorting system, which can classify garbage into the four categories: recyclable, harmful, kitchen and other garbage, and put the classified garbage into a corresponding garbage can. The system is in a standby state and does not perform any action when there is no garbage. For classifying garbage accurately, we customize an appropriative garbage sorting model based on a binarized neural network (BNN), and it is a small and low complexity network structure deployed to resource-limited embedded devices. To improve the energy efficiency of the system, we design a coprocessor dedicated to accelerating BNN on FPGA. The experiment shows that the classification accuracy of the model reaches 96.8% in the dataset we collected. The coprocessor is 5x better performance and 285x higher energy efficiency than NVIDIA GTX 1060. The system can sort 20 kinds of garbage, and the average execution time of the system is 1.3 seconds.