Image-Based Weld Joint Type Classification Using Bag of Visual Words

Increased shortage of skilled workers and increased demand for goods tend to strain manufacturing activity. This issue, along with a poor, hazardous working environment, furthers the need to robotize the activity. Welding, one of the major manufacturing processes, has witnessed automation in the last two decades. Welding using robots is mainly accomplished in ‘teach and playback’ mode. It necessitates reconfiguration every time the robot engages in a new task. Knowing the weld joint beforehand allows the programmer to set relevant parameters in advance. Hence, this study aims to solve the issue by proposing an alternate way to automatically recognize weld joint types. This paper suggests an effective way to classify the weld joint type using the image processing and feature extraction technique. The method works in two stages: features extraction and bag of visual words (BoVW) model building. First, image processing algorithms are used to condition the greyscale image. Image conditioning involves noise removal using a contrast-limited adaptive histogram equalization (CLAHE) and enhancement to improve the image’s contrast. Then SURF features of processed images are extracted and input into a support vector machine (SVM)-based bag of visual words classifier for classification. The method is capable of recognizing five types of weld joints. The bag of features strategy combined with SVM yields 97% accuracy.