Efficient Defect Detection from Consecutive Monocular Images by Deep Learning

This paper deals with the problem of detecting surface cavities on metal castings that must be distinguished from their lookalike non-defective substances embedding on the surface. As general detection methods working on the basis of single 2D appearance and patterns are easily fooled by their similar 2D appearance and less suitable, the 3D geometry, e.g. depth, of the scene must also be considered. However, ground-truth depth maps are formidable and cumbersome to obtain, even with the help of modern depth-sensing devices. We present an automated framework for effectively detecting surface cavities from only monocular vision by leveraging consecutive images from different views. We first employ an object detector to localize regions of interest (ROIs) that are candidates for defect detection, including cavities and cavity lookalikes, intuitively identical to the way human vision operates by focusing on the most important regions. We then predict and extract the pixel-wise depths of the detected ROIs with a depth estimator, which is trained in an unsupervised manner from sequences of 3 consecutive monocular images taken at different camera poses based on their relative depth consistency. The object detector and the depth estimator jointly track the ROIs over a number of images from continuous views for thorough examination. Experimental results with a real-world dataset demonstrate that our method is able to identify and characterize cavity defects into deeper regions for ease of detection.