A TinyML based Residual Binarized Neural Network for real-time Image Classification

Abstract

Image processing is an important requirement in IoT applications such as robotics, augmented reality, computer vision, Industry 4.0 etc. The capabilities of IoT devices for image processing are limited to sensing the environment, processing and communicating the results. Tiny machine learning (TinyML) is a new paradigm that takes advantage of the IoT to deploy deep learning models to perform complex tasks in resource constrained embedded devices. Image classification is an important task in IoT to interpret images of a particular scene or class. Currently, this task is performed in embedded devices using Binarized Neural Networks (BNNs), which can be converted to a set of weights using a one-hot encoding process. These networks integrated with hardware accelerators can be trained to perform image processing tasks in real-time. This paper proposes a BNN for image classification based on residual learning paradigm, called Tiny-BNN which exploits the skip connections to reduce information loss, and improve the training time and accuracy. Experimental results show that the model achieves a classification accuracy of 90.1 % and 91.6% on the on CIFAR-10 and MNIST datasets respectively.