Hidden Dot Patterns Recognition using CNNs on Raspberry Pi Zero W

Abstract

A talking pen is the evolutional way of learning especially for children because it is fun and interactive. Commercial talking pen use a small infrared camera to capture the hidden dot patterns that hidden around a book's page. Specific sounds are generated according to the location of the pen on the pages. This image processing for recognition technique is complex and need calculation power because it had to calculate relative location and orientation of dots in a group. Although, the talking pen is successful product but it is not flexible to add new pattern or uses different set of patterns. In this research, the new approach for hidden dot pattern recognition is proposed. Convolutional neural networks or CNNs is selected as the recognition software. The system is built on Raspberry Pi Zero W hardware with Raspbian operating system and TensorFlow platform. MobileNet is used in the research since it small and effective for limited resources platform. MobileNet is trained by using the infrared images that captured by infrared camera module that connected to Raspberry Pi. Result shows that the accuracy of the proposed method is good enough and the response is fast enough to be implemented into the real-time commercial product and has ability to expand to any other kind of patterns without limit because the learned network can be updated easily without alter recognition software.