Deep Learning for Accurate Corner Detection in Computer Vision-Based Inspection

Abstract

This paper describes application of deep learning for accurate detection of corner points in images and its application for an inspection system developed for the worker training and assessment. In our local built and construction industry, workers need to be certified for their technical skills through a training and assessment process. Assessment involves trainees to understand a task given with a technical drawing, e.g. electrical wiring and trunking wall assembly, and implement it accurately in a given period of time. Typically experts manually/visually evaluate the finished assembly and decide if it’s done correctly. In this study, we employed computer vision techniques for the assessment process in order to reduce significant man hour. Computer vision based system measures dimensions, orientation and position of the wall assembly and produces a report accordingly. However, analysis depends on accurate detection of the objects and their corner points which are used as reference points for measurements. Corner detection is widely used in image processing and there are numerous algorithms available in the literature. Conventional algorithms are founded upon pixel based operations and they return many redundant or false corner points. In this study, we employed a hybrid approach using deep learning and Minimum Eigen value corner detection for this purpose and achieved highly accurate corner detection. This subsequently improved the reliability of the inspection system.