Academy & Industry Research Collaboration Center (AIRCC)

Volume 12, Number 06, March 2022

Deep Learning Frameworks Evaluation for Image Classification on Resource Constrained Device


Mathieu Febvay and Ahmed Bounekkar, Université de Lyon, France


Each new generation of smartphone gains capabilities that increase performance and power efficiency allowing us to use them for increasingly complex calculations such as Deep Learning. This paper implemented four Android deep learning inference frameworks (TFLite, MNN, NCNN and PyTorch) to evaluate the most recent generation of System On a Chip (SoC) Samsung Exynos 2100, Qualcomm Snapdragon 865+ and 865. Our work focused on image classification task using five state-of-the-art models. The 50 000 images of the ImageNet 2012 validation subset were inferred. Latency and accuracy with various scenarios like CPU, OpenCL, Vulkan with and without multi-threading were measured. Power efficiency and realworld use-case were evaluated from these results as we run the same experiment on the device's camera stream until they consumed 3% of their battery. Our results show that low-level software optimizations, image pre-processing algorithms, conversion process and cooling design have an impact on latency, accuracy and energy efficiency.


Deep Learning, On-device inference, Image classification, Mobile, Quantized Models.