Banana Ripeness
Classification.

A four-class image classification study using MobileNetV2 transfer learning, with an experiment to measure how CLAHE preprocessing affects model performance.

View documentation Read the method

Overview of the banana ripeness classification research

Python & TensorFlowMobileNetV2CLAHE comparison

Research documentation

From dataset to error analysis.

Four concise visuals explain the research question, experiment, results, and the classes that remained difficult.

Project overview with the four banana ripeness classes — 01 / OVERVIEWProblem and dataset

Experiment design for MobileNetV2 transfer learning and CLAHE preprocessing — 02 / METHODExperiment design

Comparison of classification results with and without CLAHE — 03 / RESULTSModel comparison

Error analysis for banana ripeness classification — 04 / ANALYSISError patterns

Research pipeline

A controlled comparison of image preprocessing.

Dataset preparation

The study began with approximately 1,800 banana images grouped into four stages: unripe, half-ripe, ripe, and overripe. Images were resized to 224 × 224 pixels, then augmented with flips and rotations to improve variation during training.

Transfer learning experiment

MobileNetV2 was used as the convolutional neural network backbone through transfer learning. Two pipelines were trained and evaluated: a baseline using the original images, and a second pipeline using CLAHE on the L channel in LAB color space.

Evaluation

The models were compared through predictions and confusion matrices. The analysis focused on overall classification behavior and the overripe class, where visual similarities made errors easier to observe.

What I built

One study, four practical stages.

The workflow covers data, preprocessing, model training, and evaluation.

01 / DATA

Prepare the dataset

Organize four ripeness labels, resize every image, and create train and evaluation inputs.

02 / PREPROCESSING

Test CLAHE

Apply local contrast enhancement on the luminance channel while preserving color information.

03 / MODEL

Fine-tune MobileNetV2

Use a pretrained CNN backbone to learn the visual differences between ripeness stages.

04 / EVALUATION

Inspect the errors

Compare confusion matrices and review which adjacent ripeness stages the model confuses.

Research stack

Tools used in the experiment.

A compact stack for reproducible computer vision work.

Python: Dataset processing, experiment orchestration, and evaluation.
TensorFlow: Model training and inference for the four-class classifier.
MobileNetV2: Pretrained convolutional backbone used for transfer learning.
OpenCV / CLAHE: Contrast preprocessing applied in LAB color space.