This project seeks to find a possible solution for the visual examination of welds, where feature extraction methods are examined and tested with two different classifiers. The idea behind the project is to investigate if visual inspection based on texture describing features, processed with a machine learning algorithm, can detect flaws and defects in a weld merely by inspecting the surface of the object.
Visual inspection is the primary way of evaluating weld seams, where construction is not critical and additional cost is the main risk [1]. Visual inspections entail manual interpretation and evaluation, which are time consuming, and the result often depends on the person assigned to the task [1], which makes automation interesting.
The project is based on other research projects regarding the visual inspection of welds and will strive to devise a solution that can detect one type of defect that is visible to the human eye.
A dataset containing images of both good and bad welds is created from weld samples produced specifically for this purpose. For preparation of the images, different image processing tools are applied in the making of the dataset. The dataset is tested on two different feature extraction methods in the search for features that best explain image textures. To test extracted features, two classification models are tested to find the most suitable, and their results are discussed. As a result of this, a machine learning algorithm is trained on data with known targets, and tests on unknown data (processed images) are performed to analyse and compare results. Several settings, both within feature extraction and classification, are trailed and results are discussed.