Swin transformer-based shape-from-focus technique for accurate 3D shape estimation

Accurately estimating the three-dimensional (3D) shape of an object from two-dimensional (2D) images is a fundamental challenge in computer vision, with widespread applications in fields such as biomedical imaging and microscopy. Shape-from-Focus (SFF) is a widely adopted technique for addressing this challenge, relying on focus measures from image series captured at varying focal lengths. However, existing SFF techniques often face limitations such as inadequate performance in low-texture regions, sensitivity to noise, and high computational costs when processing high-resolution image series, hindering their effectiveness in real-world applications. To address these limitations, this study proposes a novel SFF technique based on the Swin Transformer, which introduces a hierarchical feature extraction framework and integrates a spatial frequency-based focus function directly into the deep feature space. This enables more robust focus-level detection and improved accuracy in 3D shape estimation for both real and simulated image series. The main contributions include multi-scale attention-based feature modeling and an enhanced sharpness evaluation strategy, yielding superior performance across various data sets. To validate the effectiveness of the proposed technique, extensive experiments are conducted using standard evaluation metrics, complemented by qualitative comparisons via reconstructed 3D shapes. The results consistently demonstrate that the Swin Transformer-based model outperforms existing SFF techniques in both quantitative and visual analyses. These findings highlight the robustness and innovation of the proposed SFF technique, establishing it as a reliable and efficient solution for 3D shape estimation.