Build notes · Goh Kun Ming
CNN Image Classification on Multi-Resolution Vegetable Dataset
AI · Public project
Apr 2025 – May 2025
Project overview
A coursework comparison of convolutional architectures, image resolutions and augmentation for vegetable classification.
A vegetable-image classification study comparing convolutional network designs across two image resolutions and augmentation settings. It combines dataset repair, repeatable preprocessing and evaluation of historical experiment results.
Inside this build
- Compared 11 vegetable classes across 23×23 and 101×101 grayscale inputs, with and without augmentation.
- Corrected validation/test directory names and misplaced carrot images using helpers with a dry-run mode.
- Explored Sequential, Functional, residual, Inception-style and depthwise-separable CNN families.
- Documented a historical Sequential search with 0.908 validation accuracy and a separate run with about 94.2% test accuracy; the results do not establish a universal architecture winner.
- Added 13 tests across Python 3.11 and 3.12, with 48% broad coverage.
- Recorded dataset split counts and missing reproducibility artifacts, including model weights and multi-seed comparisons.