Dual-level Fuzzy Learning with Patch Guidance for Image Ordinal Regression

TL;DR

DFPG tackles image ordinal regression with only image-level labels by emulating patch-focused human reasoning. It combines an Offline Patch Annotator with Adjacent Category Mixup, a Dual-level Fuzzy Learning framework that fuzzifies features at patch and channel levels, and an Online Co-teaching strategy for robust patch filtering. Across Adience, Aesthetics, and Diabetic Retinopathy, DFPG yields consistent gains over state-of-the-art methods and notably improves minority-class performance by leveraging adjacent-category information. This approach enhances robustness to label ambiguity and ordinality, enabling more reliable ordinal grading in real-world vision tasks, with code available for reproduction $\implies$ DFPG advances patch-guided, fuzzy-ordinal learning in computer vision.

Abstract

Ordinal regression bridges regression and classification by assigning objects to ordered classes. While human experts rely on discriminative patch-level features for decisions, current approaches are limited by the availability of only image-level ordinal labels, overlooking fine-grained patch-level characteristics. In this paper, we propose a Dual-level Fuzzy Learning with Patch Guidance framework, named DFPG that learns precise feature-based grading boundaries from ambiguous ordinal labels, with patch-level supervision. Specifically, we propose patch-labeling and filtering strategies to enable the model to focus on patch-level features exclusively with only image-level ordinal labels available. We further design a dual-level fuzzy learning module, which leverages fuzzy logic to quantitatively capture and handle label ambiguity from both patch-wise and channel-wise perspectives. Extensive experiments on various image ordinal regression datasets demonstrate the superiority of our proposed method, further confirming its ability in distinguishing samples from difficult-to-classify categories. The code is available at https://github.com/ZJUMAI/DFPG-ord.

Figures (5)

• Figure 1: Illustrating some influencing factors in human decision-making processes. (a) Doctors focus on specific lesion areas in the DR grading scenario. (b) Evaluators may increase aesthetic scores based on certain regions (e.g., white rectangle) while reducing scores due to blurred areas (e.g., yellow rectangle).
• Figure 2: An overview of our DFPG approach. (a) The Patch Annotator module for generating patch-level pseudo-labels. An adjacent category sampling scheme is adopted to preserve ordinal information inherent in the augmented features, thereby enhancing the model's discriminability on samples of adjacent categories. (b) The Dual-level Fuzzy Learning module, in which multiple Gaussian membership functions are used to introduce fuzziness into the precise image representations, effectively capturing the ambiguity in feature-label associations specific to ordinal regression tasks. (c) The overall Co-teaching Strategy of our model, which incorporates a patch-level optimization objective through an unreliable patch filtering method based on the generated pseudo-labels.
• Figure 3: Detailed performance for each category on the DR and Adience datasets. We show two evaluation metrics, Recall and F1-score. The star symbol indicates the minority category.
• Figure 4: Datasets visualization. Examples of each of the ordinal categories in three datasets.
• Figure 5: Detailed performance for each category on the Aesthetics datasets. We show two evaluation metrics, Recall and F1-score. The star symbol indicates the minority category.