Perspective transformation based-initial value estimation for the speckle control points matching in an out-of-focus camera calibration using a synthetic speckle pattern.

Despite camera calibration methods using regular planar chessboard or circular marker array calibration targets having been widely used, the control point extraction accuracy is low if the image is defocused or if the noise level is high. Due to the noise robustness of digital image correlation (DIC) in speckle image matching, random speckle pattern is a better choice for camera calibration than chessboard or circular markers, if the imaging quality is low. The foremost process of this method is to conduct speckle control points matching DIC, where the initial value must be estimated close to the true value. It is challenging to provide accurate initial values for DIC if the difference of physical pixel scale is large between the reference image and the target image or if the target image is out-of-focus. To solve this problem, this work presents an efficient initial value estimation method for speckle control points matching using DIC, based on perspective transformation. Firstly, the four pairs of corners of the speckle regions in the reference image and target image are detected. Secondly, the target image is transformed to a new image that has the considerable size of pixel scale with the reference image, then four neighborhood points of the control point in the reference image and the corresponding points in the transformed new image are matched coarsely by fixed subset searching. Lastly, the matched points in the transformed target image are transformed back to the origin target image by the inverse perspective transformation matrix, then the initial value for DIC can be estimated by the matched four pairs of neighborhood points. Experiment results confirm the higher calibration accuracy delivered by the proposed method, rather than that of the chessboard or the circular marker array. Measurement precision is higher than the speckle pattern calibration method that uses SIFT-based initial value estimation.