Design of deep convolutional neural network architectures for automated feature extraction in industrial inspection Daniel Weimer a,c , Bernd Scholz-Reiter (1) b, *, Moshe Shpitalni (1) c a BIBA – Bremer Institut fu ¨r Produktion und Logistik GmbH, University of Bremen, Bremen, Germany b University of Bremen, Bremen, Germany c Department of Mechanical Engineering, Technion – Israel Institute of Technology, Haifa, Israel 1. Introduction Optical Quality Control (OQC) and machine vision are vital processes in manufacturing to satisfy customer requirements, which in return will establish sales continuity of the company. To have satisfactory products, one essential step in OQC is to ensure that the product is visually free of imperfections or defects, among other requirements. In many manufacturing industries, human inspection is still a critical element in the process. Although visual inspection is a trivial task for humans to solve, in fast-paced modern industry such a task can be highly repetitive and mind- numbing, which potentially leads to human error due to fatigue. However reliable this may be, OQC performance uncertainty due to human error could be very expensive for the company. Moreover, with increasing production volume, performance of human-based OQC does not scale well as it is constrained with low inspection frequency, not to mention the cost. For these reasons, automated defect detection naturally emerges as a solution to this problem. The aim of the defect detection process is to segment a possible defective area from the background and classify it in predefined defect categories. In a controlled environment, characterized by stable lighting conditions, simple thresholding techniques are often sufficient to segment defects from the background. However, such methods are no longer applicable when dealing with surfaces with strong or complex textures or noisy sensor data. In these much more challenging applications, more elaborate methods are required to ensure stable and reliable defect detection results. As described by Xie [1] and Neogi et al. [2], existing methods can roughly be divided into four main categories: statistical, structural, filter based, and model based. In industrial defect detection systems, intensive studies have been performed in order to hand- craft the optimum feature representation of the data for the given problem. Pernkopf and O’Leary [3] for example, performed a comparative study on feature selection from an exhaustive list of different feature encodings. Jiang et al. demonstrated the power of filter-based methods, namely three different wavelet formulations, for surface analysis [4]. While these methods usually yield satisfactory results for known problems, they cannot always be applied to new or different problems set because each problem has its own characteristics that only responds to certain kind of feature extractor. Thus, it is common practice in industrial environments that new feature has to be manually engineered when a new set of problems arises. Additionally, surface defects can occur in arbitrary size, shape and orientation. Therefore, standard feature descriptors for defect description often lead to insufficient classification results [5]. Given these considerations, this contribution investigates an alternative approach, namely Convolutional Neural Networks (CNN), in order to overcome the difficulties of redefining manually a specific feature representation for every new inspection problem. A CNN consists of an arbitrary initialized set of filters, where the goal of a supervised learning procedure is to learn the best filters for a given problem. CNN automatically generate meaningful features for a specific task in an evolutionary way directly from huge amounts of raw data with minimal human interaction. In this contribution, we CIRP Annals - Manufacturing Technology xxx (2016) xxx–xxx A R T I C L E I N F O Keywords: Quality assurance Artificial intelligence Deep machine learning A B S T R A C T Fast and reliable industrial inspection is a main challenge in manufacturing scenarios. However, the defect detection performance is heavily dependent on manually defined features for defect representation. In this contribution, we investigate a new paradigm from machine learning, namely deep machine learning by examining design configurations of deep Convolutional Neural Networks (CNN) and the impact of different hyper-parameter settings towards the accuracy of defect detection results. In contrast to manually designed image processing solutions, deep CNN automatically generate powerful features by hierarchical learning strategies from massive amounts of training data with a minimum of human interaction or expert process knowledge. An application of the proposed method demonstrates excellent defect detection results with low false alarm rates. ß 2016 CIRP. * Corresponding author. E-mail address: bsr@biba.uni-bremen.de (B. Scholz-Reiter). G Model CIRP-1481; No. of Pages 4 Please cite this article in press as: Weimer D, et al. Design of deep convolutional neural network architectures for automated feature extraction in industrial inspection. CIRP Annals - Manufacturing Technology (2016), http://dx.doi.org/10.1016/j.cirp.2016.04.072 Contents lists available at ScienceDirect CIRP Annals - Manufacturing Technology journal homepage: http://ees.elsevier.com/cirp/default.asp http://dx.doi.org/10.1016/j.cirp.2016.04.072 0007-8506/ß 2016 CIRP.