Topological anchoring and adaptive penalties: the hybrid awred architecture for detect recognition in contaminated visual data

DOI: 10.31673/2412-9070.2026.027603

Abstract

In practical computer vision systems, the assumption of fully clean training data is often violated. Under structural contamination, anomalous visual samples can be partially absorbed into the latent space of normal patterns, which reduces the stability of unsupervised anomaly detection models and complicates reliable defect recognition. This issue becomes especially important when the concentration of visual distortions is high enough to affect the optimization process itself. To address this problem, the paper considers the Hybrid AWRED v3 methodology, representing the third generation of the proposed approach for defect recognition in contaminated visual data. The architecture combines topological anchoring, adaptive weighting of reconstruction errors, staged penalization, and an external Bayesian optimization loop for hyperparameter tuning.
The experimental study was conducted on the modified MNIST-C benchmark under two contamination regimes, 5% and 20%, and compared the proposed method with several baseline models, including standard autoencoders, denoising autoencoders, Deep SVDD, and DAGMM. In the moderate contamination scenario (P = 0.05), Hybrid AWRED v3 achieved the highest AUC-ROC = 0.874 ± 0.028, indicating competitive anomaly ranking quality under limited poisoning. In the severe contamination scenario (P = 0.20), the proposed architecture obtained the highest Recall = 0.880 ± 0.005 among all compared methods, outperforming DAGMM (0.397 ± 0.039) and Deep SVDD (0.584 ± 0.163) in terms of detection completeness. At the same time, the highest mean AUC-ROC in this regime was observed for DAE (0.785 ± 0.008), which suggests that the compared models differ in their sensitivity to ranking quality and to the minimization of missed defects.
Overall, the obtained results indicate that Hybrid AWRED v3 is particularly effective in scenarios where reducing the number of missed anomalies is more important than maximizing a single aggregate metric. Its main advantage in the present study is therefore most directly supported by recall under severe contamination, while broader claims of superiority across all evaluation criteria would require additional evidence. This makes the method a promising candidate for machine vision applications operating in noisy or weakly controlled environments, where contamination of the training sample cannot be excluded in advance.

Keywords: anomaly detection, data poisoning, computer vision, Hybrid AWRED, topological anchoring, Bayesian optimization, staged learning (Curriculum Learning).