TopoLeaf: A Zero-Shot Visual Anomaly Detection Framework via Stable Feature Dimensionality and Local Persistent Homology for Self-Organizing Agricultural Cyber-Physical Systems

Abstract

Abstract Zero-shot visual anomaly detection in complex textured domains remains a fundamental challenge for building adaptive, self-organizing cyber-physical systems. Conventional deep learning approaches often rely on closed-set assumptions, require prohibitive pixel-level annotation costs, and suffer severe performance degradation under cross-domain shifts---limiting their deployability in real-world agricultural CPS where novel disease types and unseen crop species continuously emerge. To address these issues, we present TopoLeaf, a training-free and annotation-free anomaly detection framework. By leveraging the robust semantic representations of foundation models (specifically DINOv2), our method introduces two complementary scoring mechanisms: a geometric anomaly score based on local KNN distance in a stability-selected feature subspace, and a topological anomaly score derived from local persistent homology. The topological score effectively captures subtle structural deviations and micro-texture mutations that geometric distances often miss. Extensive experiments on cross-species plant disease benchmarks (3,100+ images across 40 source--target pairs) demonstrate that TopoLeaf achieves highly competitive and structurally robust zero-shot performance, providing a robust perception layer for closed-loop agricultural cyber-physical systems that must maintain diagnostic stability under previously unseen perturbations. Under well-aligned domains, our geometric score achieves near-perfect detection (e.g., 0.994 AUROC on Strawberry). The method exhibits informative failure modes on structurally isolated domains such as Corn (0.169 AUROC), revealing fundamental structural properties of the foundation model's feature manifold. Furthermore, the topological score demonstrates structural complementarity, achieving 0.542 AUROC on the challenging Corn-to-Apple pair where geometric scoring degenerates to 0.221. Module ablation studies confirm that stability-based dimensionality selection consistently improves cross-domain generalization.

Plant phenotyping relevance

植物病害の視覚的異常（植物の病徴・状態）を推定する新規画像解析手法を開発し、複数種の病害ベンチマークで検証しているため、植物フェノタイピング手法が中心である。

Code and data availability

The paper publicly releases its complete TopoLeaf source code (implementation, baselines, evaluation scripts) under the MIT License on GitHub, and all experimental image data derives from the publicly available PlantVillage dataset, which is the leaf-image input used for the paper's anomaly-detection phenotyping and is