A number recognition system with memory optimized convolutional neural network for smart metering devices

This paper presents a number recognition system based on a memory-optimized convolutional neural network for smart metering devices. Smart metering is one of the fastest growing applications for wireless sensor networks. Wireless sensor nodes are in general battery powered, and thus are often constrained by limited memory size and computation power. Due to the memory constraint, general architectures of convolutional neural networks are not suitable for smart metering devices. It is also challenging to recognize the number images of smart metering devices, since the numbers are rolling on mechanical wheels. We propose a memory-optimized architecture of convolutional neural network (MO-CNN) well suited to smart metering devices with a tight memory constraint. We implemented the proposed MO-CNN in a C program and conducted experiments with various rolling number images captured using real water meters. The proposed architecture demonstrate 100% recognition rate under the light condition of 2 ∼ 150 Lux, while it reduces the memory size by 30 times compared with the conventional CNN architecture.