Noise Hierarchies in Fine-Grained Image Recognition: Pixel Bias, Label Corruption, and Purification Strategies
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Fine-grained image recognition has small inter-class differences, which can easily lead to label noise during annotation. However, most existing denoising studies primarily focus on ImageNet/CIFAR, lacking systematic evaluation for fine-grained scenarios. This paper uses CUB-200-2011 as the test set and designs a 2×2 factorial experiment to independently control pixel-level and label-level noise. The results show that under an 18% pollution intensity, label flipping leads to an 8-16% decline in Top-1 performance, while Gaussian pixel noise causes a decline of less than 2%; when both coexist, there is no significant additional degradation, which once again verifies that "label quality is the primary bottleneck for fine-grained tasks." Subsequently, under the same noise setting, this paper conducts a horizontal comparison between two general strategies for addressing label noise—Robust Loss GCE and Semi-Supervised Cleaning DivideMix: Under an 18% random flip condition, they respectively elevate the Top-1 accuracy from a baseline of 53.14% to 68.54% and 71.02%, achieving relative gains of 15.4 and 17.9 percentage points over Cross-Entropy (CE), with gaps of less than 5 parts per thousand (ppt) from the clean upper bound of 66.09%. Experiments confirm that introducing only a general loss function or a semi-supervised framework can achieve considerable gains in fine-grained bird classification tasks, without the need for specially designing a bird-specific denoising module. This paper provides a one-click reproduction script and CSV noisy labels, which can serve as a directly usable benchmark for subsequent research.
Fields of Study & Subjects
Machine Learning and Data Classification
Advanced Neural Network Applications
Domain Adaptation and Few-Shot Learning
Publication details
- Journal / source
- IEEE International Conference on Power, Electronics and Green Energy (ICPEGE)
- Pages
- 37–41
- Publication date
- 28 January 2026
- Language
- EN