Deep Tiny Quantization for Fish-Eye Distorted Object Classification

Abstract

Tiny machine learning has proven its capabilities and applicability in several research fields such as IoT and Automotive applications. The introduction of the deeply quantized neural network has been a game changer as it allowed to reduce dramatically the memory footprint. The challenge is to achieve a marginal accuracy drop low enough while quantizing 32 bits floating point neural networks. In case of mice studies, by acquiring the appropriate images per each use case, with the neural networks proposed by this work, it is possible to classify the objects inside the mice’ cages and if they drink or not. The outcomes are important to indicate the health status of the rodents. In that context, pBottleNet, pFoodNet, pCageNet have been introduced to classify the presence of the bottle, the food level and the presence of the cage while pDrinkingNet was designed to identify if the rodent was drinking when the bottle was present in the cage. The accuracies of the above cited four deeply quantized neural networks were between 95.70% and 99.9%. The entire process, from the image capture to the inference’s execution, have been deployed on microcontrollers. The design of the networks, therefore, shall respect the memory constraints of the STM32H7 and of the STM32L4 microcontrollers in which the models have been analyzed and tested. The inference times on the STM32H7 for each pico model were 1. 912ms, 12.579ms, 2. 263ms and 2. 264ms respectively.