A wavelet-guided transformer approach for autofocus in brightfield biological microscopy.

Autofocus plays a crucial role in Biological Microscopy by ensuring image clarity and improving operational efficiency. However, mainstream brightfield biological microscopes still rely on conventional autofocus methods, which suffer from poor real-time performance and high sensitivity to noise, limiting their applicability in time-critical scenarios. To address these challenges, we propose a Wavelet-Guided Transformer Network (WGT-Net) that enables fast and accurate autofocus prediction from a single blurred image. WGT-Net integrates three key design elements: the use of wavelet transform to construct multi-scale blurred features and perform downsampling; a Transformer module that captures global-local dependencies across multi-scale image features; a Gaussian soft labeling strategy that models the optimal focus position as a probability distribution to handle uncertainty. Experiments conducted on a locally collected dataset demonstrate that WGT-Net achieves a mean absolute error (MAE) of 0.0869 and a root mean square error (RMSE) of 0.101, achieving 28.69% and 32.39% reductions in MAE and RMSE, respectively, compared with state-of-the-art methods, and completing predictions within milliseconds. These results demonstrate that WGT-Net significantly improves both prediction accuracy and real-time performance, highlighting its suitability for real-time, high-throughput Brightfield Biological Microscopy applications.