How Do Grayscale Images Affect Visual Anomaly Detection?

1. Introduction: Why grayscale pictures would possibly have an effect on anomaly detection.
2. Anomaly detection, grayscale pictures: Fast recap on the 2 major topics mentioned on this article.
3. Experiment setting: What and the way we evaluate.
4. Efficiency outcomes: How grayscale pictures have an effect on mannequin efficiency.
5. Pace outcomes: How grayscale pictures have an effect on inference pace.
6. Conclusion

1. Introduction

On this article, we’ll discover how grayscale pictures have an effect on the efficiency of anomaly detection fashions and study how this selection influences inference pace.

In laptop imaginative and prescient, it’s properly established that fine-tuning pre-trained classification fashions on grayscale pictures can lead to degraded performance. However what about anomaly detection models? These fashions don’t require fine-tuning, however they use pre-trained classification fashions akin to WideResNet or EfficientNet as characteristic extractors. This raises an vital query: do these characteristic extractors produce much less related options when utilized to a grayscale picture?

This query isn’t just educational, however one with real-world implications for anybody engaged on automating industrial visible inspection in manufacturing. For instance, you would possibly end up questioning if a shade digital camera is important or if a less expensive grayscale one might be adequate. Or you may have issues relating to the inference pace and need to use any alternative to extend it.

2. Anomaly detection, grayscale pictures

If you’re already conversant in each anomaly detection in laptop imaginative and prescient and the fundamentals of digital picture illustration, be at liberty to skip this part. In any other case, it gives a short overview and hyperlinks for additional exploration.

Anomaly detection

In laptop imaginative and prescient, anomaly detection is a fast-evolving discipline inside deep studying that focuses on figuring out uncommon patterns in pictures. Sometimes, these fashions are skilled utilizing solely pictures with out defects, permitting the mannequin to be taught what “regular” appears like. Throughout inference, the mannequin can detect pictures that deviate from this discovered illustration as irregular. Such anomalies usually correspond to numerous defects that will seem in a manufacturing surroundings however weren’t seen throughout coaching. For a extra detailed introduction, see this link.

Grayscale pictures

For people, shade and grayscale pictures look fairly comparable (except for the dearth of shade). However for computer systems, a picture is an array of numbers, so it turns into a little bit bit extra sophisticated. A grayscale picture is a two-dimensional array of numbers, usually starting from 0 to 255, the place every worth represents the depth of a pixel, with 0 being black and 255 being white.

In distinction, shade pictures are usually composed of three such separate grayscale pictures (known as channels) stacked collectively to type a three-dimensional array. Every channel (red, green, and blue) describes the depth of the respective shade, and its mixture creates a shade picture. You may be taught extra about this here.

3. Experiment setting

Fashions

We’ll use 4 state-of-the-art anomaly detection fashions: PatchCore, Reverse Distillation, FastFlow, and GLASS. These fashions characterize several types of anomaly detection algorithms and, on the similar time, they’re extensively utilized in sensible functions resulting from quick coaching and inference pace. The primary three fashions use the implementation from the Anomalib library, for GLASS we make use of the official implementation.

Dataset

For our experiments, we use the VisA dataset with 12 classes of objects, which gives a wide range of pictures and has no color-dependent defects.

Metrics

We’ll use image-level AUROC to see if the entire picture was categorized appropriately with out the necessity to choose a specific threshold, and pixel-level AUPRO, which exhibits how good we’re at localizing faulty areas within the picture. Pace might be evaluated utilizing the frames-per-second (FPS) metric. For all metrics, larger values correspond to raised outcomes.

Grayscale conversion

To make a picture grayscale, we are going to use torchvision transforms.

For one channel, we additionally modify characteristic extractors utilizing the in_chans parameter within the timm library.

The code for adapting Anomalib to make use of one channel is offered here.

4. Efficiency outcomes

RGB

These are common pictures with crimson, blue, and inexperienced channels.

Grayscale, three channels

Pictures had been transformed to grayscale utilizing torchvision remodel Grayscale with three channels.

Grayscale, one channel

Pictures had been transformed to grayscale utilizing the identical torchvision remodel Grayscale with one channel.

Comparability

We are able to see that PatchCore and Reverse Distillation have shut outcomes throughout all three experiments for each picture and pixel-level metrics. FastFlow turns into considerably worse, and GLASS turns into noticeably worse. Outcomes are averaged throughout the 12 classes of objects within the VisA dataset.

What about outcomes per class of objects? Perhaps a few of them carry out worse than others, and a few higher, inflicting the common outcomes to seem the identical? Right here is the visualization of outcomes for PatchCore throughout all three experiments exhibiting that outcomes are fairly secure inside classes as properly.

The identical visualization for GLASS exhibits that some classes may be barely higher whereas some may be strongly worse. Nonetheless, this isn’t essentially brought on by grayscale transformation solely; a few of it may be common end result fluctuations resulting from how the mannequin is skilled. Averaged outcomes present a transparent tendency that for this mannequin, RGB pictures produce the very best end result, grayscale with three channels considerably worse, and grayscale with one channel the worst end result.

Bonus

How do outcomes change per class? It’s doable that some classes are merely higher fitted to RGB or grayscale pictures, even when there are not any color-dependent defects.

Right here is the visualization of the distinction between RGB and grayscale with one channel for all of the fashions. We are able to see that solely pipe_fryum class turns into barely (or strongly) worse for each mannequin. The remainder of the classes turn into worse or higher, relying on the mannequin.

Further bonus

If you’re all for how this pipe_fryum appears, listed below are a few examples with GLASS mannequin predictions.

5. Pace outcomes

The variety of channels impacts solely the primary layer of the mannequin, the remainder stays unchanged. The pace enchancment appears to be negligible, highlighting how the primary layer characteristic extraction is only a small a part of the calculations carried out by the fashions. GLASS exhibits a considerably noticeable enchancment, however on the similar time, it exhibits the worst metrics decline, so it requires warning if you wish to pace it up by switching to at least one channel.

6. Conclusion

So how does utilizing grayscale pictures have an effect on visible anomaly detection? It relies upon, however RGB appears to be the safer guess. The affect varies relying on the mannequin and information. PatchCore and Reverse Distillation usually deal with grayscale inputs properly, however you should be extra cautious with FastFlow and particularly GLASS, which exhibits some pace enchancment but additionally probably the most vital drop in efficiency metrics. If you wish to use grayscale enter, you should check and evaluate it with RGB in your particular information.

The jupyter pocket book with the Anomalib code: link.

Observe creator on LinkedIn for extra on industrial visible anomaly detection.

References

1. C. Hughes, Transfer Learning on Greyscale Images: How to Fine-Tune Pretrained Models (2022), towardsdatascience.com

2. S. Wehkamp, A practical guide to image-based anomaly detection using Anomalib (2022), weblog.ml6.eu

3. A. Baitieva, Y. Bouaouni, A. Briot, D. Ameln, S. Khalfaoui, and S. Akcay. Beyond Academic Benchmarks: Critical Analysis and Best Practices for Visual Industrial Anomaly Detection (2025), CVPR Workshop on Visible Anomaly and Novelty Detection (VAND)

4. Y. Zou, J. Jeong, L. Pemula, D. Zhang, and O. Dabeer, SPot-the-Difference Self-Supervised Pre-training for Anomaly Detection and Segmentation (2022), ECCV

5. S. Akcay, D. Ameln, A. Vaidya, B. Lakshmanan, N. Ahuja, and U. Genc, Anomalib (2022), ICIP