Benchmark for data-centric image classification an‎d noisy label estimation

1/28/2026 12:00:00 AM

Artificial Intelligence (AI) has so far primarily focused on building better models rather than creating better datasets. We have witnessed significant advancements in Machine Learning (ML) model architectures, driven by datasets that serve as benchmarks for eva‎luating model performance. In modern machine learning, high-quality data is crucial, an‎d nearly all learning models improve as data quality increases. This research addresses the problem of image classification in the presence of noisy labels, which is one of the major challenges in deep learning. Noisy labels can significantly degrade the performance of learning models, leading to reduced accuracy an‎d efficiency. In this study, we first reviewed methods to han‎dle noise in labels an‎d proposed a new framework to enhance the learning process. This framework includes data refinement an‎d interactive learning between models to identify an‎d correct noisy labels. Additionally, two complementary methods were integrated into this framework: threshold annealing, which dynamically an‎d progressively adjusts threshold values to improve data refinement, an‎d self-ensembling, which enhances model stability an‎d accuracy. These methods were tested on the CIFAR dataset, an‎d the results demonstrated that combining these approaches effectively reduced the impact of noise on the data an‎d improved model performance. Finally, recommendations for further development an‎d enhancement of these approaches are provided, which can pave the way for future research in learning with noisy data.

طبقه‌بندي تصاوير , يادگيري عميق , برچسب داراي نويز , پالايش داده , آستانه‌گذاري تدريجي , خود-تجمعي

Image classification , deep learning , noisy labels , data refinement , threshold annealing , self-ensembling

Farnaz Hasani

Morteza Analoui