Out-of-focus blur in microscopy data causes both standardized analysis methods and domain experts to misinterpret the microstructure. As a result, images affected by out-of-focus blur must either be reacquired or discarded. In both cases, this leads to an unwanted loss of valuable time and resources. In this research study, we propose an automatic in-situ focus correction for light-optical microscopy in material science. During the image acquisition, a trained deep learning (DL) model estimates the out-of-focus blur per pixel of a single image as a deviation from the sharp focus plane in micrometers. The average value of the estimated focus deviations in an image is used for a hardware readjustment of the microscope objective in the Z-plane. Consequently, the light microscope with the trained DL model is able to assess and correct the focus deviation in an image and capture a sharp image without the use of a Z-stack. In addition, the trained DL model can be integrated into an advanced automatic image processing or applied to images that have already been captured. This enables blurred image regions to be automatically identified and processed by either discarding them or restoring them using image deblurring methods. It also provides domain experts with an overview of the image sharpness. We show that the error of the focus deviation estimated by the trained DL model is within the depth of field of the microscope objectives used. Furthermore, we provide information on the training dataset and the architecture used.