Answers are grounded in this paper and cite their sources.
• A method to clean training image datasets affected by the presence of noise • Identifies measurement noise (e.g. corrupted or irrelevant samples) • Detects label noise (samples with incorrect labels) • Separates both noise types to remove unusable images and flag mislabels for review • Improves classification accuracy in real-world datasets such as insect images AI model performance requires high-quality training data, yet in agriculture, collecting large, clean datasets is challenging due to time-consuming acquisition and biological variability, making data cleaning fundamental. Traditional cleaning often focuses on measurement noise, corrupted inputs like artifacts or non-target objects that deviate from the valid data distribution and are therefore treated as outliers. Equally important is label noise, which arises when valid inputs are mislabelled and thus appear as anomalies in the feature-label relationship rather than as statistical outliers. Although both noise types have been widely studied, they are typically handled independently. As they typically co-occur, there is a demand for a unified strategy. We propose CHIMERA (Cleaning Heterogeneous Image Datasets from Measurement and Label Noise for Robust Classification Accuracy), a framework that detects and separates measurement and label noise in image classification, discarding the former and flagging the latter for relabelling. CHIMERA begins by fine-tuning a pretrained network on the noisy samples to generate predicted labels, assumed to better reflect ground truth. Based on these labels, samples are grouped and processed via feature extraction and outlier detection. Detected outliers are deemed measurement noise, while inliers with mismatched predicted and original labels are flagged as label noise. Using an insect classification task for pest monitoring, we demonstrate how CHIMERA improves dataset quality and classification performance, achieving 90.52% accuracy, compared to 87.93% on the noisy dataset. By disentangling measurement and label noise, CHIMERA provides a practical approach to clean image datasets, enhancing the robustness of models in resource-constrained agricultural settings.
No takes yet. Share an insight, caveat, or question.
Esmeral et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: