PatchFlow: Leveraging a Flow-Based Model with Patch Features

Die casting plays a crucial role across various industries due to its ability to craft intricate shapes with high precision and smooth surfaces. However, surface defects remain a major issue that impedes die casting quality control. Recently, computer vision techniques have been explored to automate and improve defect detection. In this work, we combine local neighbor-aware patch features with a normalizing flow model and bridge the gap between the generic pretrained feature extractor and industrial product images by introducing an adapter module to increase the efficiency and accuracy of automated anomaly detection. Compared to state-of-the-art methods, our approach reduces the error rate by 20% on the MVTec AD dataset, achieving an image-level AUROC of 99.28%. Our approach has also enhanced performance on the VisA dataset , achieving an image-level AUROC of 96.48%. Compared to the state-of-the-art models, this represents a 28.2% reduction in error. Additionally, experiments on a proprietary die casting dataset yield an accuracy of 95.77% for anomaly detection, without requiring any anomalous samples for training. Our method illustrates the potential of leveraging computer vision and deep learning techniques to advance inspection capabilities for the die casting industry

💡 Research Summary

PatchFlow introduces a novel unsupervised visual anomaly detection framework tailored for the die‑casting industry, but its contributions extend to any industrial quality‑control scenario where only normal samples are available for training. The method builds on four key ideas. First, it extracts multi‑level feature maps from a frozen, ImageNet‑pretrained CNN backbone (e.g., ResNet‑50, EfficientNet) and constructs “patch‑aware” descriptors by aggregating the local neighbourhood of each spatial location across several hierarchical layers. By defining a patch size ∫ and a neighbourhood set N(h,w)∫, the authors compute a patch feature pᵢʲ(N) through a simple aggregation function (e.g., average pooling). This yields a set of patch‑level descriptors that preserve fine‑grained defect cues often lost in global embeddings. Second, to bridge the domain gap between the generic pretraining data and the specific distribution of industrial die‑casting images, a lightweight adaptation module A is inserted. It consists of a single fully‑connected layer followed by batch normalization, reducing the dimensionality of the concatenated multi‑scale patch vector fᵢ,∫ and aligning its distribution toward a standard normal. This adaptation dramatically improves the stability of the subsequent normalizing‑flow (NF) stage while adding negligible computational overhead. Third, the NF component is streamlined by employing a bottleneck coupling block. The input patch vector is first compressed to a lower‑dimensional bottleneck, then transformed with the usual affine scale‑and‑shift operations, and finally expanded back. This design retains expressive power but reduces the number of parameters and the cost of Jacobian determinant computation, addressing the classic NF drawback of high FLOPs. Fourth, anomaly scoring is performed directly on the log‑likelihood produced by the flow. Image‑level scores are the average log‑likelihood over all patches, while pixel‑level heatmaps are obtained by reshaping the per‑patch likelihoods and normalizing them, providing intuitive visual explanations without any pixel‑level supervision. The entire pipeline—feature extraction, patch aggregation, adaptation, flow mapping—is trained end‑to‑end on normal images only; no anomalous samples are required. Empirical evaluation on three benchmarks demonstrates the effectiveness of the approach. On the widely used MVTec AD dataset, PatchFlow achieves an image‑level AUROC of 99.28 %, reducing the error rate by roughly 20 % compared with the previous state‑of‑the‑art PatchCore method. On the VisA dataset, it reaches 96.48 % AUROC, a 28.2 % error reduction over competing NF‑based models. Moreover, on a proprietary die‑casting dataset containing real‑world defects such as blisters, peeling, and pits, the model attains 95.77 % detection accuracy despite being trained solely on defect‑free samples. In addition to accuracy gains, the proposed bottleneck flow and lightweight adapter cut the total parameter count and FLOPs by about one‑third relative to conventional NF architectures, enabling near‑real‑time inference on commodity GPUs. The authors also discuss limitations: the choice of patch size ∫ and adapter dimensionality C_g can be dataset‑dependent, requiring careful hyper‑parameter tuning; the current design focuses on 2‑D images, leaving 3‑D scans or video streams for future work; and more sophisticated domain‑adaptation techniques (e.g., adversarial style transfer) could further close the gap between pretraining and industrial domains. In summary, PatchFlow successfully merges local patch‑level context with efficient normalizing‑flow density estimation, delivering a high‑performance, low‑cost solution for unsupervised defect detection that is directly applicable to die‑casting quality control and beyond.