Work

Dual-Manifold Gradient Rectification: Reconciling Topology and Texture in Industrial Anomaly Synthesis

Year: 2026
Type: article
Abstract: The reliability of automated industrial defect detection systems is fundamentally constrained by the long-tail distribution and scarcity of anomalous samples. This limitation is further intensified by the topological–textural conflict: rare defects often exhibit high-frequency, stochastic morphologies that are topologically incompatible with the highly structured and repetitive background manifolds of industrial components. To address this challenge, this paper proposes the Dual-Manifold Gradient Rectification (DMGR) framework. Unlike prior approaches that rely on heuristic blending or monolithic inpainting, DMGR reformulates defect synthesis as a deterministic trajectory rectification process. Specifically, a global diffusion model establishes a Base Structural Field, while a specialized local defect model provides a Rectification Field governed by manifold consistency constraints. To ensure physical plausibility, a Physics-Grounded Spectral-Spatial Alignment mechanism is introduced, which dynamically injects high-frequency defect details into the global trajectory via uncertainty-adaptive spatial gating. Furthermore, an Asymmetric Collaborative Evolution strategy is employed to coordinate the interaction between Structural Anchoring and Textural Resonance, thereby effectively harmonizing global topology with local defect fidelity. Experimental results on a PCB defect dataset demonstrate that DMGR achieves favorable semantic coherence and spectral authenticity. Augmenting the training set with the synthesized samples yields a 10.3 percent relative improvement in the detector’s mean Average Precision evaluated over multiple intersection-over-union thresholds. Notably, the proposed framework exhibits strong scalability, maintaining consistent performance gains even when the training data are augmented by a factor of six, which establishes DMGR as a practical and effective solution for addressing complex industrial anomalies. The code and data are available at https://github.com/weihang-luo/DMGR
Source: IEEE Access
Language: English
Open Access status: gold
Estimated APC paid: $2,075
OpenAlex ID: W7128301259
FWCI: 4.222
Cites: 40
Cited by: 2
Related to: 10