CornViT

A Multi-Stage Convolutional Vision Transformer Framework for Corn Kernel Analysis

Overview

Three-stage hierarchical classification pipeline for automated corn kernel quality assessment:

• Stage 1: Purity detection (Pure vs Impure)
• Stage 2: Shape classification (Flat vs Round)
• Stage 3: Embryo orientation (Up vs Down)

Architecture

• Model: CvT-13 (384×384) with ImageNet-22k pretraining
• Framework: PyTorch + Microsoft CvT
• Test Accuracy: 93.8% (Stage 1), 94.1% (Stage 2), 91.1% (Stage 3)

Setup

# Clone repository
git clone https://github.com/microsoft/CvT.git

# Install dependencies
pip install -r requirements.txt

Training

Each stage has independent training scripts:

python stage1/train_cvt13.py  # Purity classification
python stage2/train_cvt13.py  # Shape classification
python stage3/train_cvt13.py  # Embryo orientation

Inference

python stage1/inference_cvt13.py
python stage2/inference_cvt13.py
python stage3/inference_cvt13.py

Baselines

ResNet50 and DenseNet121 baselines available in baselines/.

Structure

├── stage1/          # Purity classification
├── stage2/          # Shape classification
├── stage3/          # Embryo orientation
└── preprocess/      # Data preprocessing scripts

Requirements

• Python 3.13+
• PyTorch 2.9+
• CUDA (optional, for GPU training)

---

Citation

If you use this code, models, or catalog in your research, please cite:

@Article{computers15010002,
  AUTHOR = {Erukude, Sai Teja and Mascarenhas, Jane and Shamir, Lior},
  TITLE = {CornViT: A Multi-Stage Convolutional Vision Transformer Framework for Hierarchical Corn Kernel Analysis},
  JOURNAL = {Computers},
  VOLUME = {15},
  YEAR = {2026},
  NUMBER = {1},
  ARTICLE-NUMBER = {2},
  URL = {https://www.mdpi.com/2073-431X/15/1/2},
  ISSN = {2073-431X},
  DOI = {10.3390/computers15010002}
}