Focus Score Prediction – Project Report

Final Results

• Test Mean Absolute Error (MAE): ~0.28
• Best Validation MAE: ~0.24
• Dataset Used: Out of ~5000 CSV entries, ~1064 valid images were available for training/testing (rest missing).

---

Training vs Validation Loss

Below is the plot of training and validation loss (MSE) over epochs.

---

Approach

1. Data Preprocessing

   • Used the provided focus_labels.csv which contained image_path and focus_score.
   • Cleaned dataset by dropping missing image paths.
   • Grouped images by their stack folder and performed GroupShuffleSplit to avoid data leakage between train, validation, and test sets.
   • Applied image augmentations (random flip, brightness, contrast) to improve generalization.

2. Model Architectures

   • Implemented two options:

     • Simple CNN: 3 Conv layers + Dense layers for baseline.
     • MobileNetV2 (Pretrained): Transfer learning with ImageNet weights, frozen base, followed by:
       • Global Average Pooling
       • Dropout (0.3)
       • Dense (64, ReLU)
       • Dense (1, regression output)

   • Final experiments used MobileNetV2 as it provided better performance.

3. Training Setup

   • Optimizer: Adam (1e-4)
   • Loss: Mean Squared Error (MSE)
   • Metrics: Mean Absolute Error (MAE)
   • Batch size: 32
   • Epochs: 25 (with Early Stopping)

   Callbacks used:

   • ModelCheckpoint (best model saved)
   • EarlyStopping (patience 7, restore best weights)
   • ReduceLROnPlateau (factor 0.5, patience 3)

4. Results

   • Achieved Validation MAE ~0.24 and Test MAE ~0.28.
   • The final best model was saved in Keras format (best_model.keras) and converted to TFLite float16 (focus_model.tflite) for efficient deployment.

---

Key Files

• train.py – Training pipeline script
• focus_labels.csv – Input CSV with labels
• outputs/ – Directory containing:
  • best_model.keras – Saved best model
  • focus_model.tflite – Converted TFLite model
  • training_plots.png – Training/Validation loss plot

---

✅ This project demonstrates an end-to-end deep learning pipeline for predicting focus score from cell stack images, with dataset preprocessing, training, evaluation, and deployment-ready model conversion.