D3R-Net: Dual-Domain Denoising Reconstruction Network for Robust Industrial Anomaly Detection

TL;DR

D3R-Net tackles the challenge of blurriness in reconstruction-based industrial anomaly detection by introducing a self-supervised healing task and a frequency-domain FFT magnitude loss. The method trains a compact autoencoder on normal data with on-the-fly synthetic corruptions and enforces dual-domain consistency to preserve fine textures, achieving improved localization (PRO AUC) while maintaining fast inference. Across 15 MVTec AD categories, D3R-Net with FFT loss shows consistent gains over MSE-based baselines and offers a practical alternative to heavier feature-embedding approaches. The work highlights the value of combining denoising-style training with spectral regularization and suggests future exploration of localized frequency analyses to further bridge the gap to large pretrained models.

Abstract

Unsupervised anomaly detection (UAD) is a key ingredient of automated visual inspection in modern manufacturing. The reconstruction-based methods appeal because they have basic architectural design and they process data quickly but they produce oversmoothed results for high-frequency details. As a result, subtle defects are partially reconstructed rather than highlighted, which limits segmentation accuracy. We build on this line of work and introduce D3R-Net, a Dual-Domain Denoising Reconstruction framework that couples a self-supervised 'healing' task with frequency-aware regularization. During training, the network receives synthetically corrupted normal images and is asked to reconstruct the clean targets, which prevents trivial identity mapping and pushes the model to learn the manifold of defect-free textures. In addition to the spatial mean squared error, we employ a Fast Fourier Transform (FFT) magnitude loss that encourages consistency in the frequency domain. The implementation also allows an optional structural similarity (SSIM) term, which we study in an ablation. On the MVTec AD Hazelnut benchmark, D3R-Net with the FFT loss improves localization consistency over a spatial-only baseline: PRO AUC increases from 0.603 to 0.687, while image-level ROC AUC remains robust. Evaluated across fifteen MVTec categories, the FFT variant raises the average pixel ROC AUC from 0.733 to 0.751 and PRO AUC from 0.417 to 0.468 compared to the MSE-only baseline, at roughly 20 FPS on a single GPU. The network is trained from scratch and uses a lightweight convolutional autoencoder backbone, providing a practical alternative to heavy pre-trained feature embedding methods.

Figures (4)

• Figure 1: Self-supervised healing task for D3R-Net (multi-category benchmark run). Each row shows a clean hazelnut image, its synthetically corrupted version, and the reconstruction produced by D3R-FFT. Corruptions include local occlusions, noise patches and blended foreign patches, and the model is trained to remove them and recover the normal texture.
• Figure 2: Image-level ROC curves on three representative categories (bottle, cable, hazelnut) from the multi-category benchmark. Each plot is taken directly from the benchmark run and compares reconstruction-based methods with feature-embedding baselines.
• Figure 3: Qualitative comparison on a Hazelnut test sample. Each panel shows the input, reconstruction (if applicable), anomaly map and overlay.
• Figure 4: D3R reconstructions and anomaly maps for an object-centric (hazelnut) and a texture-centric (leather) category. The same dual-domain training behaves differently depending on whether the model primarily reconstructs an object or a surface.