| --- |
| language: en |
| license: apache-2.0 |
| library_name: keras |
| tags: |
| - image-classification |
| - tensorflow |
| - efficientnet |
| - computer-vision |
| - cats-vs-dogs |
| metrics: |
| - accuracy |
| - auc |
| - precision |
| - recall |
| - f1 |
| pipeline_tag: image-classification |
| --- |
| |
| # Pet Classification with EfficientNetB0 |
|
|
| This repository contains a high-performance deep learning model designed to classify images into two categories: **Cats** and **Dogs**. The model leverages the **EfficientNetB0** architecture, utilizing Transfer Learning and specialized Fine-Tuning to achieve professional-grade metrics. |
|
|
| ## Model Performance |
|
|
| Evaluated on a balanced test set of **5,000 images**, the model demonstrates exceptional stability and discriminative power: |
|
|
| | Metric | Score | |
| | :--- | :--- | |
| | **Test Accuracy** | **97.48%** | |
| | **AUC Score** | **0.9974** | |
| | **Precision** | **96.77%** | |
| | **Recall** | **0.9824** | |
| | **F1-Score** | **0.9750** | |
|
|
| ### Confusion Matrix Highlights |
| * **Total Correct:** 4,874 / 5,000 images. |
| * **Sensitivity:** High recall for 'Dog' class (0.9824), ensuring minimal false negatives. |
| * **Confidence:** Average Loss of **0.0651**, indicating high certainty in classifications. |
|
|
| ## Architecture & Training Strategy |
|
|
| The model uses a multi-stage training pipeline to maximize the features learned from the ImageNet-pre-trained EfficientNetB0 base. |
|
|
| ### 1. Model Structure |
| * **Base:** EfficientNetB0 (Functional) |
| * **Pooling:** GlobalAveragePooling2D |
| * **Normalization:** BatchNormalization for training stability. |
| * **Dense Layers:** 256 units (ReLU) followed by a 2-unit Softmax output. |
| * **Regularization:** Dropout (0.4) to ensure high generalization and prevent overfitting. |
|
|
| ### 2. Training Phases |
| * **Phase 1 (Transfer Learning):** The base model was frozen, and only the custom classification head was trained (Learning Rate: 1e-3). |
| * **Phase 2 (Fine-Tuning):** The top 40 layers of the EfficientNet base were unfrozen and trained with a reduced learning rate (1e-4) to refine high-level feature detection. |
|
|
| ## How to Use |
|
|
| To use this model locally with the `.keras` file: |
|
|
| ```python |
| import tensorflow as tf |
| from tensorflow.keras.applications.efficientnet import preprocess_input |
| import cv2 |
| import numpy as np |
| |
| # Load model |
| model = tf.keras.models.load_model('efficientnetb0_pet_classifier_finetuned.keras') |
| |
| def predict(img_path): |
| img = cv2.imread(img_path) |
| img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) |
| img = cv2.resize(img, (224, 224)) |
| img = preprocess_input(np.expand_dims(img, axis=0)) |
| |
| preds = model.predict(img) |
| return "Dog" if np.argmax(preds) == 1 else "Cat" |
