Approximated Memory With IQA-Based Accuracy Estimation for Animal Classifiers: A Case Study

Image classification is facilitated by the proliferation of image data from various Internet of Things (IoT) and smart devices. However, the on-site employment of deep learning (DL)-based classifiers is hindered by notable energy consumption in storing those image data. Hence, this work explores the feasibility of approximated memory for animal classification, where approximation reduces image data for optimised memory usage and corresponding energy efficiency. Three different approximation algorithms are compared for five DL models to identify the optimum approach. Additionally, a mathematical model is proposed for estimating the performance of approximated memory-incorporated classifiers, facilitating application-wise optimum approximation case selection. Experimental results indicate rounding-based approximation as the optimum approach while addressing the superiority of EfficientNet-b0 with approximated memory for animal classification. Also, this work highlights 50% to 62.5% image data reduction for optimised memory usage while maintaining 96% to 99% of original accuracy for EfficientNet-b0.