Image Classifier for Recyclable Paper Classification
Project Overview:
Certain papers, such as those with wax or plastic coatings, or papers that are dirty or wet, may not be recyclable due to contamination. The goal of this project is to solve this problem with an image classifier combined with an object detection model. The integrated model will determine whether paper items are recyclable or not. The object detection component will locate paper items within an image, and the subsequent classifier will provide a binary classification output, indicating whether the paper item is recyclable or non-recyclable.
Dataset:
While numerous resources are available for recyclable papers, there are fewer resources for non-recyclable papers. To fix this imbalance and create a well-rounded dataset, we plan to download non-recyclable images, including items such as dirty papers, pizza boxes, coffee cups, and paper plates, from the internet. This approach aims to create a more balanced and better dataset for our project.
Datasets:
Model Development:
Our Model has 2 parts:
- Object Detection
- Image Classifier
For the Object Detection part, we plan to utilize YOLOv5 or YOLOv8. Additionally, we will label the train data mostly ourselves.
For image classifier part we will use PyTorch with different models like ResNet, MobileNet, EfficentNet. But however creating a robust dataset is more important for us. So we will spend most of our time creating a better dataset. Especially for non-recyclable class.
Related Work:
[1]Classification and recycling of recyclable garbage based on deep learning — ScienceDirect
This study made by Yujin Chen, Anneng Luo, et al. Uses ShuffleNet v2 and the depth-separable convolution method to create lightweight YOLOv5s for classifying and positioning recyclable waste. This study achieves 94% accuracy.
[2][PyTorch] Garbage Classification (~95% accuracy) | Kaggle
This model uses ResNet50 and does a solid job classifying garbage with an impressive 95% accuracy. Worth mentioning, our dataset overlaps with the one they used in this project.
The research focuses on optimizing the fully-connected layer of DenseNet121 using a genetic algorithm (GA) alongside data augmentation to improve waste classification accuracy. The optimized DenseNet121 achieved a high accuracy of 99.6% on the TrashNet dataset. Our dataset overlaps with TrashNet dataset as well.
[4]ThungYang-ClassificationOfTrashForRecyclabilityStatus-report.pdf (stanford.edu)
This study uses support vector machines (SVM) and a convolutional neural network (CNN) with scale-invariant feature transform (SIFT) features to sort items as recyclable or not. Although there were difficulties in finding the best settings for the models, the SVM worked better than the CNN in tests using a dataset of 400–500 images for each of the six waste categories.
