Image Edge-Based CNN Architecture for Detecting Mosaic Augmentation in Datasets: A Novel Approach for Assuring Data Quality

This research uncovers a pioneering method for distinguishing mosaic augmentation in datasets, an area yet to be thoroughly investigated. Utilizing conventional visual preprocessing tactics, such as Gaussian blurring and Canny Edge Detection, together with a Convolutional Neural Network (CNN), we develop a strategy for detecting mosaic enhancements in datasets. This methodology achieved an remarkable accuracy rate. A custom-created dataset was used to verify our proposed model, which exhibited outstanding precision in classification tasks, therefore showcasing its capacity to assess whether a dataset has been subjected to mosaic augmentation. Our technique offers substantial enhancements to dataset validation procedures, enhancing the accuracy and trustworthiness of ensuing data-centric models or systems. The paper elucidates our methodological underpinnings, detailing the experimentation setup, and presents robust results validating our approach's efficacy. Specifically, it achieved an impressive precision of 94.84%, recall rate of 0.96, and an AUC of 0.98 on a proprietary dataset. These findings pledge potential for future applications in governing image data quality, persistently guaranteeing the genuineness and reliability of data across various AI and Machine Learning pursuits.

I. Introduction

Data augmentation techniques have always been a key aspect of modern deep learning paradigms for their underlying capability to increase the training data variety and hence contribute to generalization and performance enhancement. Renowned data augmentation methods include data transformations, noisy labels, CutMix [1] and Mosaic augmentation, first introduced in YOLOv4 [2].