INT8 Quantisation for Cassava Leaf Disease Classification on Raspberry Pi

Tosho Abdulahi AbdulRahman, Roseline Oluwaseun Ogundokun, Pius Adewale Owolawi, Rotimi-Williams Bello, Topside Ehleketani Mathonsi
Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments, PMLR 319:167-178, 2026.

Abstract

We investigate the impact of post-training INT8 quantisation on a CNN trained for the Kaggle Cassava Leaf Disease Classification task (5 classes). Comparing FP32 baseline versus statically quantised INT8 model, the FP32 model achieved 81.5% accuracy (F1 = 0.676, 6.23 MB, 3.04 ms CPU latency), whereas the INT8 model shrank to 1.87 MB but accuracy dropped to 11.9% (58 ms on Raspberry Pi 4). These results reveal a drastic size-accuracy trade-off, highlighting pitfalls and practical lessons for deploying CNNs on edge devices in resource-constrained settings.

Cite this Paper

BibTeX
@InProceedings{pmlr-v319-abdulrahman26a, title = {{INT8} Quantisation for Cassava Leaf Disease Classification on {Raspberry Pi}}, author = {AbdulRahman, Tosho Abdulahi and Ogundokun, Roseline Oluwaseun and Owolawi, Pius Adewale and Bello, Rotimi-Williams and Mathonsi, Topside Ehleketani}, booktitle = {Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments}, pages = {167--178}, year = {2026}, editor = {Folorunso, Sakinat and Ogundokun, Roseline and Oladipo, Francisca}, volume = {319}, series = {Proceedings of Machine Learning Research}, month = {11--14 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v319/main/assets/abdulrahman26a/abdulrahman26a.pdf}, url = {https://proceedings.mlr.press/v319/abdulrahman26a.html}, abstract = {We investigate the impact of post-training INT8 quantisation on a CNN trained for the Kaggle Cassava Leaf Disease Classification task (5 classes). Comparing FP32 baseline versus statically quantised INT8 model, the FP32 model achieved 81.5% accuracy (F1 = 0.676, 6.23 MB, 3.04 ms CPU latency), whereas the INT8 model shrank to 1.87 MB but accuracy dropped to 11.9% (58 ms on Raspberry Pi 4). These results reveal a drastic size-accuracy trade-off, highlighting pitfalls and practical lessons for deploying CNNs on edge devices in resource-constrained settings.} }
Endnote
%0 Conference Paper %T INT8 Quantisation for Cassava Leaf Disease Classification on Raspberry Pi %A Tosho Abdulahi AbdulRahman %A Roseline Oluwaseun Ogundokun %A Pius Adewale Owolawi %A Rotimi-Williams Bello %A Topside Ehleketani Mathonsi %B Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments %C Proceedings of Machine Learning Research %D 2026 %E Sakinat Folorunso %E Roseline Ogundokun %E Francisca Oladipo %F pmlr-v319-abdulrahman26a %I PMLR %P 167--178 %U https://proceedings.mlr.press/v319/abdulrahman26a.html %V 319 %X We investigate the impact of post-training INT8 quantisation on a CNN trained for the Kaggle Cassava Leaf Disease Classification task (5 classes). Comparing FP32 baseline versus statically quantised INT8 model, the FP32 model achieved 81.5% accuracy (F1 = 0.676, 6.23 MB, 3.04 ms CPU latency), whereas the INT8 model shrank to 1.87 MB but accuracy dropped to 11.9% (58 ms on Raspberry Pi 4). These results reveal a drastic size-accuracy trade-off, highlighting pitfalls and practical lessons for deploying CNNs on edge devices in resource-constrained settings.
APA
AbdulRahman, T.A., Ogundokun, R.O., Owolawi, P.A., Bello, R. & Mathonsi, T.E.. (2026). INT8 Quantisation for Cassava Leaf Disease Classification on Raspberry Pi. Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments, in Proceedings of Machine Learning Research 319:167-178 Available from https://proceedings.mlr.press/v319/abdulrahman26a.html.

Related Material