🤖 machine learning

PatchFlow: Leveraging a Flow-Based Model with Patch Features

This paper introduces PatchFlow, a novel anomaly detection framework that integrates local neighbor-aware patch features with a normalizing flow model and an adapter module to bridge the gap between generic pretrained extractors and industrial die casting images, achieving state-of-the-art performance on MVTec AD, VisA, and proprietary datasets without requiring anomalous training samples.

Original authors: Boxiang Zhang, Baijian Yang, Xiaoming Wang, Corey Vian

Published 2026-02-06

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Boxiang Zhang, Baijian Yang, Xiaoming Wang, Corey Vian

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a quality inspector at a factory that makes metal parts using a process called die casting. Think of this like pouring molten chocolate into a fancy mold to make perfect candies. Usually, the candies come out smooth and shiny. But sometimes, they have tiny bubbles, scratches, or dents. Finding these flaws by looking at them with your eyes is slow, tiring, and humans make mistakes.

The authors of this paper, PatchFlow, built a "super-spy" computer program to find these invisible flaws automatically. Here is how they did it, explained simply:

The Problem: The "Out-of-Context" Spy

Usually, computer programs that look for defects are trained on millions of everyday photos (like cats, cars, and landscapes) found on the internet. Imagine hiring a security guard who has only ever seen pictures of parks to guard a factory. Even though the guard is smart, they don't know what a factory floor looks like. They might get confused by the specific lighting or textures of metal parts.

The Solution: PatchFlow's Three-Step Trick

The authors created a system that acts like a detective who learns to spot the "weird" stuff without ever seeing a "weird" example during training. They only show the computer pictures of perfect metal parts.

1. The "Patch" Strategy (Looking at the Neighborhood)
Instead of looking at the whole metal part as one giant picture, the computer chops the image into tiny little squares, called patches.

The Analogy: Imagine looking at a brick wall. If you zoom in on just one brick, it might look normal. But if you look at that brick and its immediate neighbors, you might see a crack running through them.
What they did: The system looks at these tiny patches and checks their "neighbors." It learns what a "normal neighborhood" of metal texture looks like.

2. The "Adapter" (The Translator)
This is their big innovation. The computer uses a pre-trained "brain" (a feature extractor) that knows about general images. But since that brain is used to seeing parks and cats, it needs a translator to understand metal parts.

The Analogy: Think of the pre-trained brain as a person who only speaks English. The factory images speak Metal. The authors built a tiny Adapter Module (a translator) that sits between them. It takes the English thoughts and instantly translates them into "Metal language" so the system understands exactly what it's looking at. This makes the system much more accurate.

3. The "Flow" (The Stretchy Rubber Sheet)
Once the system understands the metal, it uses something called a Normalizing Flow.

The Analogy: Imagine you have a sheet of rubber with a perfect grid drawn on it. You stretch and twist this rubber sheet to match the shape of a perfect metal part. Because you know exactly how the rubber was stretched for a perfect part, you can spot a flaw immediately. If a new metal part comes in and the rubber sheet doesn't stretch the right way (because there's a dent or a scratch), the system knows, "Hey, this doesn't fit the pattern!"
The Result: The system calculates exactly how "weird" a patch is. If it's too weird, it's a defect.

What Did They Achieve?

The authors tested this "super-spy" on three different challenges:

The "Textbook" Test (MVTec AD Dataset): They tested it on a standard collection of images used by scientists worldwide.
- Result: It was incredibly accurate, getting 99.28% right. This was a 20% improvement over the previous best method. It's like going from a B+ student to an A+ student.
The "Hard" Test (VisA Dataset): They tested it on a harder set of images with complex objects.
- Result: It got 96.48% right, beating the previous best by 28%.
The "Real World" Test (Die Casting): They tested it on actual metal valve parts from a factory (Stellantis).
- Result: They showed the computer only pictures of perfect valves. Then, they asked it to find the bad ones. It found 95.77% of the defects correctly, even though it had never seen a defective valve before!

Why Does This Matter?

The paper claims this method is faster and more efficient than older, heavier models. It doesn't need to see thousands of broken parts to learn what a broken part looks like; it just needs to learn what a perfect part looks like and then spot anything that doesn't fit.

In short, PatchFlow is like a highly trained security guard who learns the "normal" look of a factory floor so well that they can instantly spot a single misplaced screw or a tiny scratch, making the manufacturing process safer and less wasteful.

Technical Summary: PatchFlow

Problem Statement
Die casting is a critical manufacturing process for producing high-precision, complex metal parts, yet it is plagued by surface defects such as blisters, peeling, and pits. These defects, often caused by contaminants, moisture, or temperature fluctuations, compromise product quality and performance. Traditional inspection methods relying on manual visual checks are labor-intensive, time-consuming, and prone to human error. While computer vision-based anomaly detection offers a solution, existing methods face specific challenges when applied to industrial settings:

Distribution Mismatch: Pretrained feature extractors (e.g., on ImageNet) often fail to align with the specific distribution of industrial product images.
Computational Complexity: Flow-based models, which offer precise probability density estimation, typically require complex structures with high computational loads (FLOPs) to map feature distributions to latent spaces.
Granularity: Some existing flow-based models focus on image-level predictions, lacking the precision required for pixel-level defect localization.

Methodology: PatchFlow
The authors propose PatchFlow, a patch-level anomaly detection framework that integrates local neighborhood-aware features with normalizing flow modeling. The architecture consists of four primary components:

Feature Extractor: A pretrained backbone (EfficientNet B5) extracts multi-level representations from input images. The model utilizes features from multiple hierarchical layers (specifically blocks 12, 19, and 35) to capture both global abstract information and local details.
Feature Aggregation: A module constructs local neighborhood-aware patch features across multiple hierarchies and scales. It combines descriptors from different layers by interpolating deeper features to match shallower ones and concatenating them. The system processes three distinct scales to generate multi-scale hierarchy patch features.
Feature Adaptation (Adapter): To address the distribution gap between generic pretrained features and industrial data, a lightweight adaptation module is introduced. This module employs a single fully-connected layer to transform the aggregated patch features, aligning them with the industrial data distribution without introducing significant trainable parameters.
Normalizing Flow: The adapted patch features are mapped bijectively to a standard normal distribution ( $Z \sim N(0, I)$ $Z \sim N (0, I)$ ) using a coupling layer structure based on RealNVP. Unlike prior works that map image-level features, PatchFlow maps patch-level features directly. This allows for precise defect localization based on the deviation of specific patches from the normal distribution.
- Efficiency: The coupling blocks utilize a bottleneck structure (a three-layer CNN with 1x1 kernels) to reduce computational complexity (FLOPs) while maintaining expressive capacity.

Training and Inference

Training: The model is trained exclusively on normal (defect-free) samples. The objective is to minimize the Kullback-Leibler (KL) divergence between the estimated feature distribution and the actual feature distribution, effectively learning the underlying distribution of normal data.
Inference: During testing, the model calculates the likelihood of new instances. An anomaly score is derived from the maximum value in the generated anomaly map, which is created by upsampling the normalized latent patch features via bilinear interpolation.

Key Contributions

PatchFlow Framework: A novel framework combining local neighborhood-aware patch features with flow-based likelihood modeling for anomaly detection.
Feature Adaptation Module: A lightweight mechanism that bridges the distribution gap between generic pretrained extractors and industrial images, improving flow stability and detection accuracy.
Efficient Bottlenecked Coupling: A design that reduces computational complexity in normalizing flows while preserving mapping capabilities, enabling deeper networks with fewer FLOPs.
Patch-Level Mapping: A strategy that maps individual patch features rather than whole images, facilitating fine-grained, pixel-level defect localization.

Experimental Results
The authors evaluated PatchFlow on three datasets: MVTec AD, VisA, and a proprietary die casting dataset.

MVTec AD Dataset: PatchFlow achieved an image-level AUROC of 99.28%, outperforming the state-of-the-art PatchCore (99.1%) and reducing the error rate by 20%.
VisA Dataset: The model achieved an image-level AUROC of 96.48%, representing a 28.2% reduction in error compared to PatchCore.
Proprietary Die Casting Dataset: Tested on valve body images with synthesized defects, the model achieved an accuracy of 95.77% (F1 Score) without requiring any anomalous samples for training. The confusion matrix showed 34/34 normal samples and 33/36 anomalous samples correctly identified.
Efficiency: The model operates at approximately 2.7 frames per second (FPS) on a single Nvidia GeForce RTX 2080Ti GPU, with total FLOPs of $1.25 \times 10^{10}$ .
Ablation Studies: Experiments indicated that using a 3-scale image pyramid yielded the best performance, and increasing the number of flow steps beyond one actually decreased performance, suggesting a single flow step is optimal for this configuration.

Significance and Claims
The paper claims that PatchFlow illustrates the potential of leveraging computer vision and deep learning to advance inspection capabilities for the die casting industry. By combining local neighbor-aware features with an efficient flow-based model and an adapter module, the method addresses the specific constraints of industrial anomaly detection: distribution mismatch and computational cost. The results demonstrate a viable path to improving quality control and reducing waste in industrial production environments, setting a new benchmark for image-level anomaly detection on standard benchmarks like MVTec AD and VisA. The authors emphasize that their approach achieves state-of-the-art performance without the need for anomalous training data, making it highly applicable to real-world scenarios where defect data is scarce or unavailable.

The Problem: The "Out-of-Context" Spy

The Solution: PatchFlow's Three-Step Trick

What Did They Achieve?

Why Does This Matter?

Technical Summary: PatchFlow

More like this