Autofocus is a critical task for achieving sharp imaging in cameras. Conventional global search methods are time-consuming and struggle to meet the demands of online high-speed imaging. A dual-image frequency-domain autofocus method was developed, in which a Frequency-domain Defocus Representation (FDR) was constructed from the radial power-spectrum ratio of two defocused images and was fused with spatial features extracted by a lightweight convolutional network. Through end-to-end single-step regression, the model directly predicted a signed defocus step that specified both the magnitude and direction of defocus. On a dataset covering various industrial surfaces, including metal, injection-molded parts, PCBs, and paper, the proposed method achieved a validation MAE of 1.60 fv. Compared with the dual-image baseline without FDR (2.41 fv), the introduction of FDR reduced the MAE by approximately 33.6%. Furthermore, compared with hill-climbing search, the proposed method required only two captured frames and achieved a focusing success rate of 93.8%. These results demonstrate its clear advantages for real-time autofocus in industrial measurement.

autofocus; dual-image input; frequency-domain defocus feature; power spectrum ratio