新IT卓越大讲堂 No.24 期
Abstract: Deep learning-based pattern recognition has achieved impressive accuracy recently. In particular, convolutional neural networks (CNNs) have been successfully applied to solving various practical problems. One such problem is image-based product defect inspection. Product quality inspection is an essential step in a production line. So far, this task has been conducted primarily by human inspectors, but the inspection results are often affected by various human factors such as experience, health conditions, attention, and so on. CNN can be useful for reducing human factors and thus can be good for quality control. In practice, however, CNN-based defect detection is still a challenging task. One reason is that a single false negative (FN) error (i.e., classifying a defect as a normal case) may be a liability for a company, and this is the crucial difference between image recognition and defect detection. To reduce the FN error rate, in this study, we investigate the effect of “rejection” for defect detection. Experimental results demonstrate that a CNN ensemble with a proper rejection rate can have a very low FN error rate and can reduce human labor significantly.
qiangfu zhao received the b.s. degree in computer science from shandong university (jinan, china) in 1982; the m. eng. degree in information engineering from toyohashi university of technology (toyohashi, japan) in 1985; and d. eng. degree in electronic engineering from tohoku university (sendai, japan), in 1988. he was an associate professor from 1991 to 1993 at beijing institute of technology; associate professor from 1993 to 1995 at tohoku university (japan); associate professor from 1995 to 1999 at the university of aizu (japan); and tenure full professor since 1999 at the university of aizu. he is the head of system intelligence laboratory; chair of the computer science division; associate editor of ieee transactions on cybernetics; associate editor of the international journal of machine learning and cybernetics (springer); and associate editor of ieee smc magazine. he is the co-chair of the technical committee on awareness computing in ieee systems, man, and cybernetics society and the task force on aware computing in ieee computational intelligence society. he has organized or co-organized several conferences, including the 19th symposium on intelligent systems (fan2009); the 2009 international workshop on aware computing (iwac2009); the 2010 international symposium on aware computing (isac2010); and the ieee international conference on awareness science and technology (icast2011 - 2019). he has published more than 200 referred journal and international conference papers related to optimal linear system design, neuro-computing, evolutionary computing, awareness computing, and machine learning.