Batik Vision

Problem

Batik motif datasets are small and imbalanced, while real photos vary in lighting and angle. A model trained from scratch easily memorises the training data.

Approach

I used MobileNetV2 as the backbone with light augmentation and class weights. I split the data by image source to prevent leakage between training and test sets, then evaluated with a confusion matrix.

Key features

• Seven motif classes with per-prediction confidence.
• Simple Grad-CAM to show the area the model used.
• Image upload interface with built-in examples.
• Per-class evaluation notes.

Results

1. 01

   88% validation accuracy on data not used during training.

2. 02

   Data leakage was found and removed after the first evaluation looked too good.

3. 03

   The model runs fast enough on CPU for a classroom demo.