From Handcrafted Focus Measurement Operators to Deep Learning: A Comprehensive Review of Shape from Focus Strategies

Shape from Focus (SFF) is a critical passive optical measurement technique used to reconstruct the 3D shape of objects from a sequence of 2D images captured at varying focal planes. This review provides a comprehensive and systematic analysis of the evolution of SFF strategies, tracing the transition from classical handcrafted focus measurement operators to contemporary deep learning-based architectures. We categorize traditional methods into spatial and transform-domain operators, evaluating their performance in terms of sharpness extraction and noise resilience. The paper further investigates the impact of convolutional neural networks and transformer-based models, which have redefined state of the art performance by learning hierarchical feature representations. In addition to algorithmic advancements, we examine the role of pre-processing, post-processing, and approximation frameworks in enhancing 3D reconstruction accuracy. The practical utility of these strategies is demonstrated across diverse high precision fields, including semiconductor inspection, digital morphology in healthcare. Lastly, we outline future research potential in autonomous 3D vision systems by discussing the unresolved issues in this field, such as the scarcity of ground truth data and the computational needs of real time processing.