A method to dynamically refocus a single image is presented; by combining deep learning-based light field synthesis with geometric structure-based circle of confusion rendering, it simulates the light field refocusing effect. In the proposed method, the depth map is estimated and converted into disparity, and then the circle of confusion diameter is measured at different depths to resample the pixels. Two neural network structures are designed, supervised by multi-views and refocused images of the light field camera. Experiments are conducted on multiple datasets and real scenes. Compared with other techniques, the results obtained using the proposed method show superior visual performance and evaluation indicators, along with an acceptable computational cost, with the peak signal-to-noise ratio and structural similarity index reaching 34.55 and 0.937, respectively.