Lightweight convolutional neural network architecture implementation using TensorFlow lite

Abstract

Recently, with the increase in the precision of convolutional neural networks (CNN) on a wide variety of classification and recognition tasks, the demand for their deployment has dramatically increased. Even the focus is on lightweight, faster, and low-power implementations. In this paper, we have implemented a CNN model onto an embedded platform, ‘Raspberry Pi 4-Model B edge computing system (RP4-BECS)’. This CNN model was initially trained and verified in MATLAB and then implemented on the Machine Learning (ML) framework to generate a TensorFlow lite (TF-lite) flat buffer format. This implementation offers a reduced size of models with good prediction accuracy and lesser inference time as compared with the available literature. We attempted three trials for all the digits from 0 to 9 to evaluate average prediction accuracy and average inference time. An average prediction accuracy of 99.32% and average inference time of 22.53 ms is achieved for the Sign Language Digits Database (SLDD). Further, an average prediction accuracy of 99.09% and average inference time of 13.28 ms is achieved for the Modified National Institute of Standards and Technology Database (MNIST). The model sizes implemented using TF-Lite are highly reduced to 1.53 MB for SLDD and 148 KB for the MNIST database. The obtained accuracy, inference time and model sizes are better than published results.