Advanced search

Adaptive visual inspection methodologies for low cost high performance systems

Grant number: 16/23410-8
Support type:Scholarships abroad - Research
Effective date (Start): February 15, 2017
Effective date (End): February 14, 2018
Field of knowledge:Engineering - Electrical Engineering
Principal Investigator:Carlos de Oliveira Affonso
Grantee:
Host: Olli Johannes Silven
Instituição-sede : Universidade Estadual Paulista (UNESP). Campus Experimental de Itapeva. Itapeva, SP, Brazil
Research place: University of Oulu, Finland  
Associated research grant:13/07375-0 - CeMEAI - Center for Mathematical Sciences Applied to Industry, AP.CEPID

Abstract

High added-value problems initially motivated to developing specialized, methodologically rigid single purpose industrial visual inspection systems. These solutions work with well dened inspection targets, and when the changes in manufacturing processes are known, such as modications to mechanical assembly. The inspection challenges of natural materials ranging from sorting of coee beans, berries, and fruits to wooden boards and veneer, tend to be much harder. Deviations between defects and sound material, as well as dierences between classes of defects, can be dicult to detect and and discriminate by human inspectors. Many manufactured materials fall between those types, for example, woven products, steel, road surfaces, etc., produced as continuous web and supposed to have uniform appearance or controlled variations. Computer vision textbooks have been found to provide for rather ineective models for solving these more general types of challenges. Even the recent advances in machine learning methods oer only a partial solution. We propose creating industrial visual inspection methodology for materials the (1) appearance of which may vary between batches, or (2) have variations that for a human are hard to classify in a consistent manner. The justications are: (1) in case of material changes, textbook methods require labor intensive material collection and retraining; and (2) error prone human involvement has negative impact on the accuracy and added value from inspection. The base methodology and design principles should be revised to adopt the high accuracy potential from supervised training and advanced classiers, exibility from parameter tuning approaches, and the speed from simple methodologies, while providing easy use and low cost. (AU)

Scientific publications
(References retrieved automatically from Web of Science and SciELO through information on FAPESP grants and their corresponding numbers as mentioned in the publications by the authors)
AFFONSO, CARLOS; DEBIASO ROSSI, ANDRE LUIS; ANTUNES VIEIRA, FABIO HENRIQUE; DE LEON FERREIRA DE CARVALHO, ANDRE CARLOS PONCE. Deep learning for biological image classification. EXPERT SYSTEMS WITH APPLICATIONS, v. 85, p. 114-122, NOV 1 2017. Web of Science Citations: 0.

Please report errors in scientific publications list by writing to: cdi@fapesp.br.