Waste Image Classification Dataset Preprocessing

Problem Statement

The dataset acquired for waste image classification contains variations in size, dimensions, and quality. To prepare the data for machine learning model training, a preprocessing step is necessary to standardize the data and address issues such as duplicates, missing values, and inconsistent labeling.

Dataset Link: https://www.kaggle.com/datasets/sumn2u/garbage-classification-v2/

Tasks

1. Data Cleaning

• Identify and remove duplicate images to eliminate redundancy.
• Detect and handle corrupted or incomplete images to ensure data integrity.
• Check for and handle any missing or mislabeled data for accurate labeling.

2. Resizing

• Standardize the size of all images to a common dimension suitable for model training.
• Choose an appropriate size that balances computational efficiency and image quality.

3. Normalization

• Normalize pixel values to a specific range, typically between 0 and 1.
• Normalize color channels to have a mean of 0 and a standard deviation of 1 for improved model convergence.

4. Augmentation

• Apply data augmentation techniques to artificially increase the dataset size.
• Ensure augmentation parameters align with the characteristics of waste images.

5. Labeling

• Validate and correct labels for each image to ensure accurate representation of classes.

6. Data Splitting

• Divide the dataset into training, validation, and test sets for robust model evaluation.

7. File Format

• Convert images to a common file format (e.g., JPEG, PNG).
• Check and handle different color modes to ensure consistency.

8. Handling Class Imbalance

• Implement techniques to address class imbalance, such as oversampling or undersampling, if necessary.

9. Metadata Extraction

• Extract and store relevant metadata (e.g., timestamps, source information) for potential analysis or model enhancement.

10. Save Preprocessed Data

• Save preprocessed images and corresponding labels in a structured format for easy access during model training.

Outcome

A well-preprocessed dataset ready for training machine learning models on waste image classification, addressing issues related to data quality, consistency, and standardization.

Problem

This project utilizes the Trashnet dataset, comprising of 2527 images across six classes: glass, paper, cardboard, plastic, metal, and trash.

• 501 glass
• 594 paper
• 403 cardboard
• 482 plastic
• 410 metal
• 137 trash

Training an effective waste classification model requires more high-quality, diverse data.

Actions

Create a separate waste dataset by collecting from different sources and use transfer learning approach to add new knowledge.

References

• Classification Model
• Dataset Analysis
• Research Paper
• Building Process