Adaptive Label Correction Boosts Ordinal Image Classification Accuracy

Global: Adaptive Label Correction Boosts Ordinal Image Classification Accuracy

Researchers publishing on arXiv in September 2025 introduced a novel data‑centric technique, ORDinal Adaptive Correction (ORDAC), to detect and amend noisy labels in ordinal image classification tasks, a problem that arises when class boundaries are ambiguous and labeling errors are common.

The Challenge of Noisy Ordinal Labels

Ordinal classification, such as estimating age or disease severity, relies on labeled data that often contain uncertainty; mislabeling can degrade model performance and reliability, especially under asymmetric Gaussian noise conditions.

ORDAC’s Label Distribution Learning Approach

ORDAC employs Label Distribution Learning to represent each sample with a probability distribution, dynamically adjusting its mean and standard deviation during training. Rather than discarding suspect samples, the method seeks to correct them, preserving the full dataset for optimal learning.

Benchmark Evaluation on Age and Disease Datasets

The authors evaluated ORDAC on the Adience age‑estimation dataset and a diabetic retinopathy severity set, injecting varying levels of synthetic noise to simulate real‑world labeling errors.

Quantitative Gains Demonstrated

Under a 40 % noise scenario on Adience, the ORDAC_R variant lowered the mean absolute error from 0.86 to 0.62 and raised the recall metric from 0.37 to 0.49, indicating a substantial improvement in predictive accuracy.

Extended Variants and Intrinsic Noise Correction

Extended versions, ORDAC_C and ORDAC_R, further enhanced results across datasets, and the approach also proved effective at correcting intrinsic noise already present in the original training data.

Implications for Future Ordinal Modeling

The findings suggest that adaptive correction via label distributions can increase robustness and accuracy of ordinal classification models, offering a practical pathway for handling noisy annotations in computer‑vision applications.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.