Machine learning system based on computer vision for the automatic inspection of magnetic particles in marine structures

Abstract

This work presents a system of supervised learning based on 
computer vision with the aim of solving the automation of 
non-destructive inspection tests based on magnetic particles. 
In this paper, three supervised learning algorithms have been 
tested: the nearest k neighbor (kNN), a Bayesian classifier 
(NBC) and the vector support machine (SVM). The developed 
system has been successfully tested on a set of images 
extracted during the inspection of magnetic particles on marine 
structures at the Navantia shipyard in Cartagena. The algorithm 
that offered the best result was the SVM with a sensitivity of 
98.6% and a specificity of 100.0% in the detection of faults 
by magnetic particles. The vector of characteristics used 
is composed of a set of 16 elements formed by geometric 
characteristics and intensity values of the RGB, HSV, and CIE L 
* a * b * color spaces. The work presents a software application 
and a hardware system that, using the SVM algorithm, is 
capable of automatically detecting defects on marine structures 
during the magnetic particle test