Week1

Plastic Waste Classification using CNN: Week 1 work for the Edunet-Shell Skills4Future AICTE Internship, including dataset download, preprocessing, and label visualization to classify plastic waste into organic and recyclable categories.

Dataset

The dataset folder contains two subfolders:

• TRAIN: For training the model.
• TEST: For testing the model.

Tools and Libraries

• Python
• TensorFlow and Keras
• OpenCV for image processing
• Matplotlib for visualization
• Pandas for data handling

Week 1 Work

• Downloaded the dataset and organized it into TRAIN and TEST folders.
• Created the file waste_classification.ipynb for model development.
• Preprocessed the dataset:
  • Loaded images using OpenCV and converted them to RGB format.
  • Created a DataFrame with images and their respective labels.
  • Visualized label distribution using pie charts.

---

Week 2 - Convolutional Neural Network (CNN) for Image Classification

In Week 2, we focused on building a CNN model using TensorFlow and Keras to classify images. The dataset consists of labeled images, and we performed data preprocessing, model training, and evaluation.

Implemented Features

• Data Preprocessing: Used ImageDataGenerator to rescale images and load them from directories.

• Model Architecture:

  • Three convolutional layers with ReLU activation and max-pooling.
  • Fully connected layers with dropout for regularization.
  • Final output layer using softmax activation for binary classification.

• Training Process:

  • Compiled the model using Adam optimizer and binary cross-entropy loss.
  • Trained for 15 epochs with a batch size of 64.
  • Validated performance using a test dataset.

• Visualization: Randomly displayed sample images with their corresponding labels.

Model Architectre

The model consists of the following layers:

• Conv2D (32 filters, 3x3 kernel, ReLU activation, MaxPooling)
• Conv2D (64 filters, 3x3 kernel, ReLU activation, MaxPooling)
• Conv2D (128 filters, 3x3 kernel, ReLU activation, MaxPooling)
• Flatten layer
• Fully Connected Layer (256 neurons, ReLU, Dropout 0.5)
• Fully Connected Layer (64 neurons, ReLU, Dropout 0.5)
• Output Layer (2 neurons, softmax activation for binary classification)

Training Details

• Loss Function: Binary Cross-Entropy
• Optimizer: Adam
• Metrics: Accuracy
• Batch Size: 64
• Epochs: 15
• Data Augmentation: Rescaling of pixel values

Results

Model training was successfully completed with validation on the test dataset. The trained model can classify images into two categories.

---

Week3

In Week 3 of my internship, I worked on evaluating the performance of our CNN model, implementing a prediction function for waste classification, and integrating a Streamlit-based web application for real-time classification. The model is trained to classify images into Recyclable or Organic Waste.

Implemented Features

• Performance Visualization:

  • Plotted training vs validation accuracy.
  • Plotted training vs validation loss.

• Prediction Function:

  • Implemented predict_fun(img) to classify images as either Recyclable or Organic Waste.
  • Uses OpenCV for image loading and preprocessing.
  • Utilizes the trained CNN model for classification.

• Testing the Model:

  • Loaded test images from the dataset.
  • Used predict_fun() to classify test images.

• Streamlit Application:

  • Developed Waste_classification.py using Streamlit for an interactive web-based waste classification system.

• Users can upload images, and the model will predict whether the waste is Recyclable or Organic.

Results & Observations

• he model performs well in classifying waste categories.
• The loss and accuracy plots indicate training stability.
• Some misclassifications suggest further tuning is required.
• The Streamlit application provides an easy-to-use interface for real-time classification.