Artificial Intelligence based Vision Inspection for Manufacturing Industries

Abstract

Nondestructive testing is the most economical method for material inspection used in science and technology industries to evaluate the properties of a material component or system without causing damage. Aside from straightforward sensory inspections, the most popular nondestructive testing methods for welds include liquid penetrant, radiography, magnetic particle, eddy current, and ultrasonic testing. Contaminants and metallurgical flaws can be introduced into the weld via the welding process and technique. A weak weld causes a welding fault, which weakens the joint. It is described as the location in the welding process that is outside the allowable tolerance. In our research, we have applied data augmentation technique to our dataset. Further, we have used Detectron2 which offers various detection and segmentation algorithms to perform instance segmentation to segment out various defects in the images of welds. We used Slicing Aided Hyper Inference to slice and overlap the images and performed large-scale instance segmentation using Mask R-CNN X101–FPN on 44,339 images. The approach results in 0.98 training accuracy of MASK- RCNN and a total training loss of 0.2. The Average Precision is calculated as 0.826.