EREL-Net: A Remedy for Industrial Bottle Defect Detection

摘要

Product defect detection is an integral part of quality control process in any manufacturing industry. In many cases, this problem is solved by a specific designed system for each type of product, which often requires parameter tuning for each product model. In this paper, we propose a generic method for defect detection that can be deployed for various kinds of products and models. We detect defects on bottle surface and classify bottles accordingly. Bottle defect detection is a challenging task due to several factors like no sufficient training data, reflective (metallic) bottle surface, and visually similar defects with design patterns on bottles. To overcome these challenges, we first use a computer vision-based region detection technique called EREL to extract multiple regions of interest from training images and thus increase the volume of training data. The extracted regions are manually labelled as defectiveon-defective. Then, we train our proposed CNN classifier to discriminate between defective and non-defective regions, based on the extracted regions and labels. Experimental results demonstrate superior performance on non-reflective bottles and acceptable performance of the proposed method with 77% accuracy on overall unseen test images, considering various kinds of bottles and challenging reflective metallic bottles. With a current modest personal computer, our method takes around 2.4 s to process an input image to generate final image with bounding boxes localizing the defects (if any).