README.md

metadata

title: Waste Classify
emoji: 📈
colorFrom: green
colorTo: red
sdk: gradio
sdk_version: 6.19.0
python_version: '3.13'
app_file: app.py
pinned: false
license: apache-2.0

Waste Classification – Organic vs Recyclable (VGG16)

This repository contains two Keras/TensorFlow models for classifying images of waste into Organic (O) and Recyclable (R) categories. Both models are based on a pre‑trained VGG16 backbone, but they differ in the training approach:

• O_R_tlearn_vgg16_final.keras – Feature‑extraction model: the VGG16 base is frozen and only the newly added dense layers are trained.
• O_R_tlearn_fine_tune_vgg16_final.keras – Fine‑tuned model: after feature extraction, the last convolutional block of VGG16 is unfrozen and the whole network is trained further, yielding slightly better performance.

Inference is performed using the fine‑tuned model (...fine_tune...), as it achieves the highest accuracy on the test set.

---

Model Performance

Model	Test Accuracy
Feature Extraction	80%
Fine‑tuned (recommended)	81%

---

Files

• O_R_tlearn_vgg16_final.keras – feature‑extraction model
• O_R_tlearn_fine_tune_vgg16_final.keras – fine‑tuned model (used for predictions)

---

How to Use

1. Load the Fine‑Tuned Model

import tensorflow as tf
import numpy as np
from tensorflow.keras.preprocessing.image import load_img, img_to_array

model = tf.keras.models.load_model("O_R_tlearn_fine_tune_vgg16_final.keras")

2. Preprocess an Image

IMG_SIZE = (150, 150)

def preprocess_image(image_path):
    img = load_img(image_path, target_size=IMG_SIZE)
    img_array = img_to_array(img)
    img_array = img_array / 255.0
    return np.expand_dims(img_array, axis=0)

3. Make a Prediction

image_path = "path/to/your/waste_image.jpg"
input_data = preprocess_image(image_path)
prediction = model.predict(input_data)

# Output class
if prediction[0][0] < 0.5:
    print("Predicted: Organic (O)")
else:
    print("Predicted: Recyclable (R)")

Note: The model outputs a single sigmoid value. Values below 0.5 are classified as Organic, above as Recyclable.

---

Training Details

• Base model: VGG16 (weights = imagenet)
• Input size: 150×150 pixels
• Optimizer: RMSprop with learning rate decay
• Loss: binary cross‑entropy
• Early stopping and model checkpointing were used to prevent overfitting

The dataset was split into training (80%) and validation (20%). Data augmentation (shifts, flips) was applied during training.

---

Citation

If you use this model, please cite this repository.

---

License

This model is released under the MIT License.