A Neural Network Approach for Image Classification of Welded Joints using Evolutionary Computing Algorithm

Abstract

In general, a variety of vision inspection system is used
to detect the surface defects such as problem of inaccuracy in images,
non-uniform illumination, noise and deficient contrast in welding
joints. In this work, a new machine vision inspection system is
introduced to inspect the quality level for imperfection of Metal Inert
Gas (MIG) welded joints. In this proposed system, images of welded
surfaces are captured through CCD camera. From these images, the
regions of interest are segmented and features using principal
component of the images (Eigen vector) are extracted. Principal
component analysis provides good dimensionality reduction than
other features. This procedure is repeated for four different types of
welding joints. Finally, welded joints are classified using Differential
Evolution Algorithm based Artificial Neural Networks (ANN). Eigen
vectors of images are considered as input of ANN and different types
of welded joints are considered as output of network. In this work,
welding standard EN25817 is considered for surface quality level for
imperfections. Differential Evolution Algorithm (DEA) based
Artificial Neural Network is population based search algorithm, which
is an improved version of genetic algorithm. It is found to be faster and
robust in optimization. The result of this proposed system is 98.15 in
overall accuracy level. This proposed system assures that convergence
rate of DEA based ANN holds goods.