Effective Waste Classification through Data Augmentation

Overview

This repository contains the implementation of our work titled "Effective Waste Classification through Data Augmentation," which explores the impact of various Data Augmentation (DA) techniques, including Image Manipulation, Image Erasing, and Diffusion Models on waste classification. We utilize DenseNet-121 to analyze performance improvements, generalization, image fidelity, and computational costs. Our study introduces a structured framework for hyperparameter tuning of DA methods and provides insights into the strategic selection of DA techniques based on their inherent trade-offs.

Key Findings

• Enhanced Model Performance: Our adaptive framework significantly improves model performance through balanced augmentation approaches.
• State-of-the-Art Techniques: Diffusion models, which are currently state-of-the-art, substantially enhance performance and generalization, despite their higher computational cost.
• Trade-offs in Data Augmentation: We discuss the trade-offs between different DA techniques, particularly focusing on the balance between performance enhancement and computational expenses.
• Generalization Across Datasets: The study highlights the importance of considering generalization to different datasets when evaluating augmentation methods.
• Iterative pipeline for Data Augmentation: We present a framework of iterating through multiple cycles of generation, evaluation through quantitative metrics such as KL divergence and adjustment, to refine the augmentation strategy leading to a more effective and adaptable dataset for downstream tasks.

Results

Overall results acorss pipelines

DA Technique	RealWaste(Val)	TrashNet(Test)
DenseNet-121 (DN)	81%	42%
DN + Image Manipulation	72%	53%
DN + Image Erasing	77%	52%
DN + Diffusion Models	75%	66%
DN + Combined DA	75%	63%

Metrics per Technique

DA Technique	Creation	Train	GG*	LPIPS
Image Manipulation	~ minutes	2.3	19%	0.61
Image Erasing	~ minutes	2.5	15%	0.51
Diffusion Models	~ days	2.6	9%	0.55
Combined DA	-	2.3	13%	0.57

(Note: GG* stands for Generalization Gap and Train is the ratio of the total training time between the baseline model and the current model)

Table 4: F1 Scores in Realwaste[Validation] per Category

DA Technique	Cardboard	Glass	Metal	Paper	Plastic	Trash
DenseNet-121 (DN)	79%	86%	83%	79%	82%	80%
DN + Image Manipulation	73%	75%	77%	74%	71%	69%
DN + Image Erasing	77%	78%	81%	78%	73%	81%
DN + Diffusion Models	74%	85%	79%	73%	71%	66%
DN + Combined DA	75%	83%	78%	78%	72%	73%

Table 5: F1 Scores in TrashNet[Test] per Category

DA Technique	Cardboard	Glass	Metal	Paper	Plastic	Trash
DenseNet-121 (DN)	45%	50%	60%	55%	40%	3%
DN + Image Manipulation	60%	65%	70%	72%	30%	7%
DN + Image Erasing	52%	62%	66%	61%	48%	9%
DN + Diffusion Models	69%	65%	71%	75%	56%	9%
DN + Combined DA	63%	65%	70%	72%	50%	11%

The iterative pipeline

Example of previous pipelines without Iterative Pipeline (Test Set)

As can be seen the iterative pipeline framework showed increased performance and can provide a further framework to apply in all Data Augmentation tasks.

DA Technique	Cardboard	Glass	Metal	Paper	Plastic	Trash	Overall
DN + Image Manipulation(1)	32%	45%	59%	51%	24%	2%	38%
DN + Image Erasing(1)	44%	36%	56%	61%	48%	9%	47%
DN + Image Erasing(2)	51%	44%	57%	55%	45%	5%	48%