Industrial image anomaly detection (IAD) is crucial for quality control, but its performance often degrades when training data contain label noise. To circumvent the reliance on potentially flawed labels, unsupervised methods that learn from the data distribution itself have become a mainstream approach. Among various unsupervised techniques, student–teacher frameworks have emerged as a highly effective paradigm. Student–Teacher Feature Pyramid Matching (STPM) is a powerful method within this paradigm, yet it is susceptible to such noise. Inspired by STPM and aiming to solve this issue, this paper introduces Meta-Learning Guided STPM (MLG-STPM), a novel framework that enhances STPM’s robustness by incorporating a guidance mechanism inspired by meta-learning. This guidance is achieved through an Evolving Meta-Set (EMS), which dynamically maintains a small high-confidence subset of training samples identified by their low disagreement between student and teacher networks. By training the student network on a combination of the current batch and the EMS, MLG-STPM effectively mitigates the impact of noisy labels without requiring an external clean dataset or complex re-weighting schemes. Comprehensive experiments on the MVTec AD and VisA benchmark datasets with synthetic label noise (0% to 20%) demonstrate that MLG-STPM significantly improves anomaly detection and localization performance compared to the original STPM, especially under higher noise conditions, and achieves competitive results against other relevant approaches.

Image steganography is often employed in information security and confidential communications, yet it typically faces challenges of imperceptibility and robustness during transmission. Meanwhile, insufficient attention has been paid to preserving the quality of the secret image after JPEG compression at the receiver, which limits the effectiveness of steganography. In this study, we propose an anti-compression attention-based diffusion pattern steganography model using GAN (ADPGAN). ADPGAN leverages dense connectivity to fuse shallow and deep image features with secret data, achieving high robustness against JPEG compression. Meanwhile, an enhanced attention module and a discriminator are employed to minimize image distortion caused by data embedding, thereby significantly improving the imperceptibility of the host image. Based on ADPGAN, we propose a novel JPEG-compression-resistant image framework that improves the quality of the recovered image by ensuring that the degradation of the reconstructed image primarily stems from sampling rather than JPEG compression. Unlike direct embedding of full-size secret images, we downsample the secret image into a secret data stream and embed it into the cover image via ADPGAN, demonstrating high distortion resistance and high-fidelity recovery of the secret image. Ablation studies validate the effectiveness of ADPGAN, achieving a 0-bit error rate (BER) under JPEG compression at a quality factor of 20, yielding an average Peak Signal-to-Noise Ratio (PSNR) of 39.70 dB for the recovered images.

Unsupervised anomaly detection methods for industrial inspection typically adopt a two-stage inference pipeline: a pixel-level anomaly map is first produced, and a binary image-level decision is then derived via adaptive thresholding. While practically effective, this thresholding stage lacks probabilistic grounding, yielding anomaly scores without calibrated confidence and offering no mechanism for uncertainty-aware decision making in safety-critical manufacturing environments. We propose LaRD (Laplace-regularized Reverse Distillation), a method that replaces heuristic thresholding with a Bayesian image-level classifier based on the last-layer Laplace approximation, applied post-hoc to a Reverse Distillation backbone without modifying its architecture or retraining procedure. Supervision is provided by feature-space pseudo-anomalies synthesised from normal training samples, and the Laplace approximation yields both a calibrated anomaly score and a posterior uncertainty estimate for each test image. Evaluated on four standard industrial benchmarks, LaRD consistently improves image-level AUROC and Average Precision over the Reverse Distillation baseline across all evaluated datasets. Beyond classification accuracy, the posterior uncertainty achieves Uncertainty AUROC above 0.93 on all four datasets and provides a principled confidence signal that is absent from conventional score-based approaches: it enables uncertainty-aware analysis of inspection outputs and supports risk-stratified decision making, for instance by flagging products classified as normal but carrying high posterior uncertainty for human review.