Self-calibrated three-dimensional digital image correlation based on deep learning

Abstract As a significant method for nondestructive testing, digital image correlation has found extensive applications across diverse engineering domains. However, its implementation in stereo-vision systems can be challenging, especially in scenarios where placing a calibration board is difficult or impractical. To address this challenge, the present study focuses on a binocular camera and proposes a deep-learning-based self-calibration method that eliminates the need for external targets while maintaining high calibration accuracy. Specifically, a two-step object detection-based localization algorithm is designed to automatically extract the semantic label positions. Based on the detected labels, extrinsic camera parameters are retrieved via a self-calibration process grounded in relative orientation principles. Multiple experimental validations were performed, revealing that the proposed method achieves precise localization of semantic labels, with the calculated camera parameters agreeing well with the classical chessboard calibration method. Additionally, the effectiveness of the proposed method is validated in concrete compression tests.