| --- |
| library_name: keras |
| pipeline_tag: image-classification |
| tags: |
| - tensorflow |
| - keras |
| - efficientnet |
| - food-waste |
| - image-classification |
| --- |
| |
| # SnapChef Edibility Classifier |
|
|
| This model classifies loose food images into two categories: |
|
|
| - Edible |
| - Inedible |
|
|
| ## Model Architecture |
|
|
| The model uses EfficientNet-B0 with transfer learning |
|
|
| The original pretrained layers were frozen during the first training stage |
|
|
| The final 20 EfficientNet-B0 layers were later fine tuned using a lower learning rate |
|
|
| ## Optimisation |
|
|
| The model was improved using: |
|
|
| - Data augmentation |
| - Dropout |
| - Early stopping |
| - Learning rate reduction |
| - Fine tuning |
| - Hard example retraining |
|
|
| Hard example retraining was used after a fresh red apple was incorrectly classified as inedible during real world testing |
|
|
| ## Performance |
|
|
| - Final validation accuracy: 96 percent |
| - Final validation loss: 0.1364 |
| - Test accuracy before hard example retraining: 93.33 percent |
|
|
| ## Labels |
|
|
| - 0 = Edible |
| - 1 = Inedible |
|
|
| ## Input |
|
|
| The model accepts RGB food images resized to 224 by 224 pixels |
|
|
| ## Intended Use |
|
|
| This model is designed for the SnapChef food waste reduction application |
|
|
| It helps users estimate whether loose food appears edible or inedible based on an uploaded image |
|
|
| ## Limitations |
|
|
| The model may make incorrect predictions for food types or visual conditions that were not sufficiently represented in the training dataset |
|
|
| Predictions should not replace proper food safety inspection or professional advice |
