Vision-Model Benchmarking for Crop Disease Classification
Hamza Mooraj
Visual Model Architecture Benchmarking for Robust Crop Disease Classification
Overview
I designed and implemented a unified evaluation framework to systematically compare CNNs, contrastive Vision–Language Models, and generative Vision–Language Models for fine-grained crop disease classification under real-world domain shift.
The goal was not just accuracy under controlled conditions, but robustness across lab and field environments.
Problem
Many computer vision systems for agriculture perform well on curated lab images but degrade significantly when deployed in uncontrolled field conditions.
This project evaluated multiple architectural paradigms under explicit Lab vs Field separation to understand performance trade-offs and robustness characteristics.
Dataset
AgriPath-LF16
111,307 images
16 crops
41 diseases
65 crop–disease pairs
Explicit Lab vs Field source separation
Examples of images of a tomato crop with bacterial spot in AgriPath-LF16. Image (a) is a lab-based image, and (b) is a field-based image
Available on HuggingFace:
Balanced subset (used in experiments):
https://huggingface.co/datasets/hamzamooraj99/AgriPath-LF16-30k
The 30k subset preserves all classes and supports fair evaluation across domain conditions.
System Design
Domain-aware dataset (65 crop–disease classes with clear lab and field split)
Unified training and evaluation pipeline across:
ResNet-50 (transfer learning baseline)
Contrastive VLMs (CLIP, SigLIP)
Generative VLMs (Qwen2.5-VL, SmolVLM)
LoRA fine-tuning for parameter-efficient adaptation
Deterministic decoding for generative models
Programmatic output parsing with invalid-generation penalties
Macro F1 as primary metric
Separate evaluation for Lab-only, Field-only, and combined regimes
Engineering & Infrastructure
Designed modular training pipelines supporting CNN, contrastive, and generative VLM backbones under a unified interface
Implemented parameter-efficient fine-tuning (LoRA) to reduce GPU memory footprint while maintaining adaptation capacity
Optimised training on RTX 4090 and A100 using mixed precision (bf16)
Built automated evaluation scripts handling multi-split benchmarking and structured output validation
Integrated experiment tracking and artifact versioning using Weights & Biases
Implemented deterministic inference configurations to eliminate stochastic evaluation variance
Technologies
PyTorch, Transformers, PEFT, Unsloth, PyTorch Lightning, Weights & Biases
Link to GitHub Repository
Like this project
Posted Feb 20, 2026
Benchmarked CNNs and Vision–Language Models under domain shift to build a robust crop disease classification pipeline across lab and field data.
Likes
0
Views
1
Timeline
Jun 25, 2025 - Ongoing
