Multi-surface hydraulic valve block technique hole plug inspection from monocular image

Abstract

We propose a visual inspection framework that can detect technique hole plugs on multiple surfaces of hydraulic valve block from monocular RGB image. Firstly, the pose of the valve block relative to the camera is estimated, and then the visible surfaces of the valve block are corrected. Finally, according to the size information, the image patch around the technique hole are classified to judge whether there is a plug. This paper uses the LINEMOD object pose estimation framework as the pose estimation backbone. We propose a special contour extraction method for the counterbore features of valve blocks. Sample points of a matched template are used to construct a nonlinear optimization problem to obtain the accuracy pose. The experimental results show that our method can accurately estimate the pose of the valve block, and at the same time can cooperate with a simple machine learning algorithm to complete the plug inspection.