Lightweight CNN Architecture for IoT: Enhancing Character Recognition in Multiple Fonts

Existing approaches to English character recognition generally ignore font differences, and those based on deep learning are often trained on only one font owing to computational constraints. To address this problem, we propose the Very Lightweight Network (VLNet), a lightweight convolutional neural network. First, we decompose the characters of a given font into their constituent strokes. Subsequently, we pass the stroke information as input to a simple convolutional neural network. Stroke-based feature extraction reduces the requirement for graphics data and training input size. The network is small and efficient; thus, it is suitable for edge computing and Internet of Things applications. In the experimental comparison with the standard character-recognition systems LeNet, AlexNet and MobileNet V3, VLNet demonstrated superior accuracy for known and unknown fonts and a reasonable prediction time per character. The results of this study were also implemented in Plustek Inc.’s Q30 network scanner, which enabled direct document content recognition and transfer to various cloud services.