Today marked another big step forward for my food image classifier.
After reaching ~41% accuracy yesterday with 36-class data I collected using Selenium,
I decided to push the dataset further — aiming for 100+ images per class.

And the result?

Accuracy has officially broken the 50% barrier – hitting 56%. 💥
That’s no longer random guessing — it’s learning.

🛠️ What I Did Today

• Performed a second round of crawling for underrepresented food categories
• Ensured 100+ images per class across all 36 classes
• Removed duplicates and restructured the dataset folders
• Re-trained the same MobileNetV2 model with the expanded dataset

📈 Results

• Accuracy improved from 41% to 56% 🚀
• This is the first time the model passed 50% accuracy
• Some classes like sweet potato and watermelon showed recall over 0.6
• Confusion matrix indicates real learning, not just guessing anymore
• Class imbalance still affects some predictions, but overall performance is far more robust

💡 Insight

“Better data beats better models — every time.”

No architectural changes were made.
The only change was more and better data, and the result was a significant accuracy jump.
It reminded me again that data engineering is just as important as modeling.

🎯 Next Steps

• Address class imbalance (e.g. class_weight='balanced', oversampling)
• Try out new architectures like EfficientNetB0 or ConvNeXt
• Visualize attention with Grad-CAM to understand model focus
• Build a quick demo app with Streamlit for sharing results

This phase taught me something critical:

You can’t tune your way out of bad data.
You have to build a solid foundation first — and that starts with the dataset.