.

ISSN 2063-5346
For urgent queries please contact : +918130348310

VISUAL DEFECT CLASSIFICATION IN STEEL SURFACES USING TRANSFORMER ARCHITECTURE

Main Article Content

K.Andal, S.Sathiya, P.Sivaraj
» doi: 10.31838/ecb/2023.12.s3.314

Abstract

Defect classification in a typical surface using automated visual inspection (AVI) tool for planar materials is an important task often implemented after flaw detection, and it serves as a necessary prerequisite for achieving the on-line quality inspection of finished goods. In the industrial environment of manufacturing flat steels, this detection and classification of defects is incredibly difficult due to different appearances, unclear intraclass and interclass variations.This study shall propose a classification approach using transformer architecture for classification of defects present in the steel surfaces. A sequence of vectors is created by dividing animage into fixed-size patches, linearly embedding each one, adding position embeddings, and then feeding the assembled vectors to a conventional Transformer encoder. The traditional method of performing classification involves including an extra learnable "classification token" in the sequence.By giving the network attention, it is possible to learn the connections between the image patches. This can be accomplished either in combination with a Convolutional Neural Network(CNN) model or by substituting a few of its elements. These network architectures can be used for image classification tasks.

Article Details