Original Research · 2026

CPU-Constrained Deep Learning
for Tomato Disease Detection

Obidur Rahman^*, Lipon Chandra Das, Arnab Aich, Abu Saiman Md Taiham, Atif Ibna Latif

Under Review, Springer Book Proceedings

Agriculture faces 40% yield loss from disease. While GPUs power modern AI, they remain inaccessible to developing regions where farmers rely on basic laptops. We benchmark ResNet-50, ConvNeXt-Tiny, and FastViT-T8 on consumer CPU hardware, identifying models that balance accuracy with real-world deployability.

Paper PDF GitHub Repo Dataset

Introduction

Prior studies (DenseNet, Inception V3) relied on NVIDIA Tesla or RTX GPUs — hardware financially inaccessible to smallholder farmers in South Asia and Africa. These farmers depend on consumer-grade laptops or mobile devices. Our work addresses this gap by identifying a model that balances accuracy and inference speed on standard CPUs, enabling accessible disease detection for the 180M+ ton global tomato market.

Method

We evaluated three architectures on an AMD Ryzen 5 5600G (6C/12T, no GPU).

ResNet-50

25.6M params · Baseline

Traditional CNN. Established stability but computationally heavy for CPU inference.

ConvNeXt-Tiny

29.0M params · Modern CNN

Transformer-inspired architecture with 7×7 kernels. Highest parameter count.

FastViT-T8

4.03M params · Hybrid

CNN-Transformer hybrid. 6× smaller. Optimized for edge inference.

Batch: 8 (Eff: 32)Optimizer: AdamWLR: 1e-4 → 5e-5Cosine DecayRandomResizedCrop + FlipInput: 224×224

PLACEHOLDER: architecture_diagram.jpg — Model architecture comparison

Dataset

PlantVillage subset — 16,012 images across 10 disease classes. Class imbalance ratio of 8.6:1 (Yellow Leaf Curl: 3,209 vs Mosaic Virus: 373). Standard 70/15/15 train/val/test split.

PLACEHOLDER: dataset_samples.jpg — Sample leaf images per class

Results

FastViT-T8 delivers the best speed-accuracy tradeoff: 99.66% accuracy at 0.022s/img (45 FPS), 57% faster than ConvNeXt-Tiny while losing only 0.22% accuracy. ConvNeXt-Tiny achieves the highest accuracy (99.88%) but at 0.051s/img.

Model	Accuracy	Precision	Recall	F1	Latency
ConvNeXt-Tiny	99.88%	0.999	0.998	0.998	0.051s
FastViT-T8	99.66%	0.997	0.996	0.996	0.022s
ResNet-50	97.69%	0.978	0.976	0.976	0.055s

PLACEHOLDER: benchmark_chart.png
Accuracy vs speed scatter plot

PLACEHOLDER: confusion_matrix.png
Confusion matrix heatmap

Limitations

Several caveats limit the generalisability of these findings:

Data leakage. Images split randomly, not by plant ID. Reported accuracies are likely upper bounds.
Single seed. Results based on seed=42. Multi-seed validation needed to confirm the 0.22% gap significance.
Lab conditions. PlantVillage is a controlled dataset. Real-world field images with complex backgrounds will degrade performance.
Overfitting signal. ConvNeXt reached 100% training accuracy vs 99.88% validation — suggests memorisation in the largest model.

Citation

If this work is useful in your research, please cite:

@inproceedings{rahman2026cpu, title={CPU-Constrained Deep Learning for Tomato Disease Detection: Traditional, Modern, and Hybrid CNN Comparison}, author={Rahman, Obidur and Das, Lipon Chandra and Aich, Arnab and Taiham, Abu Saiman Md and Latif, Atif Ibna}, booktitle={Springer Book Proceedings}, year={2026}, note={Under Review}, }