Cheese Image Classification with Machine Learning

UW–Eau Claire • Research Assistant to Dr. Xiang Ma • Summer 2025

What: Built a CNN in PyTorch to classify cheese types; created a reproducible, cleaned dataset and training pipeline.
My role: Data collection & cleaning, automation scripts, GPU training on HPC, evaluation, and documentation.
Impact: Working model, published cleaned dataset/code, reusable pipeline for future lab projects.

I researched how convolutional neural networks (CNNs) can classify cheese types from images. I focused on data quality and reproducibility: collecting and cleaning data, building an automated pipeline, and training/evaluating models with GPU acceleration. The project contributes a maintainable baseline for future computer vision work in food technology.

4 Months

Timeline

PyTorch

DL Framework

My Responsibilities

Collected images from Kaggle and other sources; organized class labels.
Wrote Python scripts (PIL + CleanVision) to detect duplicates, blurry/corrupt, and mislabeled images.
Integrated Google Gemini API to flag questionable images and reduce dataset noise.
Produced raw → filtered → cleaned dataset versions with hashes/manifests for reproducibility.
Submitted/monitored GPU jobs on the university HPC (logging accuracy/loss and resource usage).
Experimented with learning rate, batch size, and epochs; documented weekly reports and a final summary.

Dataset & Cleaning

Aggregated images across multiple sources. Automated quality checks with CleanVision and PIL (duplicates, blur, corrupt files, label anomalies). Used Gemini for heuristic triage of edge cases. Generated train/val/test splits with class stratification.

Preprocessing & Augmentation

Resize, center-crop, normalize; light aug (flip/rotate/color jitter) to improve generalization while preserving class-distinctive features.

Model & Training

Baseline CNN (cross-entropy) trained on GPU with LR scheduling and early stopping. Iterated hyperparameters (LR, batch size, epochs) and compared runs via consistent seeds and logs.

Evaluation & Analysis

Reported accuracy and macro-F1; inspected per-class confusion matrices and failure cases; adjusted sampling/weights to address class imbalance.

Languages

Python

Framework

PyTorch

Libraries

NumPy, Pandas, Matplotlib, PIL, CleanVision

AI APIs

Google Gemini (image triage)

Platform

University HPC (GPU), Google Colab

Ops

Job submission, logging, run reproducibility (seeds/config)

Outputs

Code repo, cleaned dataset, training pipeline

Docs

Weekly reports + final write-up

What Was Accomplished

Built a working image classification model for cheese types with a reproducible training setup.
Produced a cleaned and labeled dataset (raw → filtered → cleaned) suitable for future research.
Automated the data cleaning pipeline for repeatable runs and faster iteration.
Documented methods, results, and recommendations for the research group.

Key Findings

Data cleaning and removal of noisy samples improved stability and accuracy more than deeper models at current scale.
Class imbalance drove the largest confusions; weighting/oversampling reduced minority-class errors.
Light augmentation boosted generalization on unseen samples without degrading class cues.

Next Steps

Benchmark transfer learning (ResNet/EfficientNet) vs. custom CNN on the cleaned dataset.
Expand classes and standardize capture conditions to further reduce visual overlap.
Harden the pipeline with experiment tracking and a model registry for future studies.

← Back to Research