Today marks the final chapter of my first full computer vision project.
After dozens of experiments, hundreds of crawled images, and countless hours of training — I’m calling it.

Not because I’ve reached the perfect model.
But because I’ve reached my deadline — and that matters too.

🔄 What I Did (Recap)

Over the past week, I tried to stretch this project in every direction:

• 📸 Crawled 200+ images per class for 36 food categories
• 🔍 Improved keyword specificity ("fruit apple" vs. "food")
• 🧹 Filtered corrupted or low-quality images
• 🧪 Trained and tuned MobileNetV2 with transfer learning
• 🛠️ Applied techniques like:
  • EarlyStopping
  • ReduceLROnPlateau
  • ModelCheckpoint
  • Basic class_weight balancing

And yes — I watched accuracy climb from 14% → 61%.
But more importantly, I watched myself improve.

💭 What I Learned

This project wasn’t about getting 95% accuracy.
It was about building the pipeline from scratch, and discovering what really moves the needle:

• 🧠 Data is everything: Doubling the dataset often gave more gain than changing the model
• 🧼 Preprocessing matters: Cleaning data beat fancy architectures
• ⚖️ Ethics isn’t optional: I consciously avoided redistributing images, and made sure to disclaim copyright concerns
• 🎯 Deadlines are real: I chose to stop experiments early — not because I ran out of ideas, but because the goal was completeness, not perfection

I now understand the full CV process better:
From crawling and cleaning → to training and tuning → to reporting and reflecting.

📌 If I Had More Time…

• Try EfficientNetB0 or ResNet50
• Fine-tune pretrained weights
• Apply Grad-CAM for interpretability
• Build a Streamlit demo for the model
• Use Test-Time Augmentation or mixup

But again, experimentation never ends.
This project does.

✅ Final Metrics

• Accuracy: 61%
• Macro F1-score: 0.60
• Best performing classes: banana, pineapple, kiwi
• Worst: capsicum, apple, corn

Model: MobileNetV2 (frozen base) + custom dense layers
Optimizer: Adam, learning rate schedule enabled

🌱 Final Reflection

You don’t have to finish everything.
You just have to finish something — and learn from it.

This was my first real dive into computer vision.
I’m walking away not just with a model, but with a mindset.

And that’s worth more than accuracy.