Advanced Machine Vision for Detecting Dents and Scratches on Metal Surfaces

Science & Technology


Introduction

In the realm of quality assurance, flexible vision technology excels at the precise detection of cosmetic flaws on surfaces, particularly beneficial during visual inspections. This article outlines a detailed process of training a machine vision system to identify scratches and dents on metal surfaces, followed by an evaluation of its performance.

Creating the Training Dataset

We start by initiating a new project and preparing the metal surface by adding some scratches for our tests. Once satisfied with this preliminary step, snapshots of the metal surface are taken, rotating it to capture the surface from multiple angles. It is crucial to include varied reflections of light on the metal's grain to build a comprehensive model. The process is then repeated using a center punch to create dents of various sizes for broader sample coverage.

Annotation Process

Upon completing the photo capturing stage, we move to annotate the images. Here, two new tag types are created: one for scratches and another for dents. Each photo is meticulously reviewed, tagging every scratch and dent, zooming in to ensure fine details are captured. This step is repeated for both types of flaws, resulting in 152 scratches and 54 dents being tagged, an optimal sample size for robust training.

Training and Prediction Settings

With the annotated dataset prepared, the project is saved, and prediction settings are adjusted to randomly scale samples and alter their contrast. This introduces greater variability, amplifying the dataset size for enhanced training accuracy. Post cloud training, the model can be downloaded and run locally.

Model Performance

The model's real-time detection capabilities are then tested by introducing new scratches and dents, demonstrating rapid identification of flaws. Alternatively, the model can function via the cloud, utilizing snapshots for flaw detection. Flexible vision technology proves adept at identifying cosmetic defects under a variety of complex conditions, ensuring high-quality standards.

Keywords

flaw detection, visual inspection, quality assurance, metal surfaces, machine vision, scales and contrast, training dataset, scratch detection, dent detection, model performance

FAQs

1. How do you start training a machine vision system for detecting flaws?

We begin by creating a new project and adding distinguishable flaws to the test metal surface, then take snapshots from various angles to capture different light reflections.

2. What is the importance of varying the metal surface angles during the training phase?

Varying the angles helps in capturing how different light reflections off the metal surface affect the visibility of flaws, leading to a more robust model.

3. How do annotations contribute to the machine vision system?

Annotations are crucial as they label the flaws in the dataset, allowing the machine vision system to learn and identify these flaws accurately.

4. Why are prediction settings adjusted to scale and change contrast randomly?

Adjusting these settings introduces greater variability in the training dataset, helping the model to generalize better and perform accurately under diverse conditions.

5. Can the trained model be run both locally and on the cloud?

Yes, the trained model can be executed on a local machine for real-time detection as well as over the cloud using snapshots.