Phase-ReNet: a phase regression network for feature detection from a defocused pattern for camera calibration

Conventional camera calibration typically requires acquiring clear and focused target images for accurate feature detection. Defocused target images may reduce the feature detection accuracy and even lead to failure in estimating camera parameters. To address this issue, this paper employs the crossed fringe as a calibration pattern and develops an effective phase regression network (Phase-ReNet) for wrapped phase calibration, from which feature points can be extracted with high precision. Unlike traditional phase-shifting methods, which require multiple patterns, our method recovers horizontal and vertical phase maps using just a single pattern, significantly improving calibration efficiency. Experimental results demonstrate that this method can achieve feature detection accuracy comparable to traditional phase-shifting methods, and the mean reprojection errors of the defocused camera are only 0.0552 pixels. These results highlight that our method is suitable for defocused camera calibration tasks.