Empirical correlations for bubble diameter and velocity are incapable of predicting the local bubble behaviors fairly because the impact of local hydrodynamics on bubbles in fluidized beds. Based on image processing, a novel bubble identification method with an adaptive threshold was proposed to distinguish and characterize bubbles in fluidized beds. The information regarding bubble properties and local hydrodynamics can thus be extracted using the big data from highly resolved simulations. Accordingly, the deep neural network was trained to accurately predict local bubble properties, where the inputs were determined by performing correlation analysis and a random forest algorithm. We found Reynolds number, voidage, and relative coordinates are the dominant factors, and a four-variable choice was demonstrated to output satisfactory performance for predicting local bubble diameter and velocity. The model was preliminarily validated by coupling with the EMMS drag into CFD codes, which showed that the accuracy of coarse-grid simulations can be significantly improved.