Deep Learning and Knowledge Distillation based Flare Detection for Camera Module Inspection

Abstract

본 연구에서는 분류 과제를 수행할 수 있는 딥 러닝 네트워크와‘지식 증류’ 기법을 바탕으로 하는 고속 고성능의 폰 카메라 플레어 검사 알고리즘을 제안하고자 한다. 딥 러닝을 이용하여 양품과 불량을 분류함에 있어서 간단한 구조의 모델로도 높은 성능을 가지도록 학습하는 데 중점을 두었다. 이를 위하여 깊은 구조를 가지는 고성능의 모델 학습 결과를 단순한 구조의 모델 학습에 지식 증류 기법을 통해 전달함으로써 높은 정확도의 검출 결과를 낼 수 있도록 하였다. 즉, 본 연구는 실제 현장에 적용 가능한 알고리즘을 연구함으로써 학문적 연구를 실용적으로 사용할 수 있게 하는 데 의의를 가진다.;During the manufacturing process of the phone, camera modules of the phone goes through a defect inspection. Currently, most inspections are performed based on manual inspection by an operator or inspection based on an image processing algorithm. But this does not guarantee the consistency of the inspection and the performance might be limited. Deep convolutional neural networks (CNNs) have been successful for visual recognition tasks. Especially combined with machine vision system, CNNs have been applied to defect inspection system in industry. Despite the high performance, CNNs require high computational cost and this becomes a big huddle to be applied for real-time applications. In this paper, we study a deep learning based camera flare detection system. We designed a method that is based on cropped feature map for flare inspection. Furthermore, we compressed the deep CNN model using knowledge distillation to design a faster inspection system. In experiments using real camera modules, the proposed method showed significantly improved speed comparing to the deep CNN method. Through the study, this paper aims at suggesting an algorithm which is not only fast but also has a high performance. Therefore the algorithm would be practical enough to be applied to real industry field.