摘要： The vertical blooming of charge-coupled device (CCD) usually occurs in silicon-based cameras, which poses great challenge to determine the M2 of near-infrared (NIR) lasers. In this paper, a new method based on deep learning technique to suppress the influence of CCD vertical blooming on M2 determination is proposed for the first time, to the best of our knowledge. Taking a step-index few-mode fiber as an example, large amounts of samples including the blooming near-field beam patterns and their corresponding M2 values are used to train the convolutional neural network (CNN), aiming at learning a fast and accurate mapping from the beam pattern to the M2 parameter. The trained CNN can then be utilized to analyse the blooming pattern recorded by CCD to determine the M2. The simulated testing results have shown that the averaged prediction error of our scheme is about 0.5% for the investigated fiber beams. As for the time cost, our trained CNN only takes about 10 ms to determine the M2 value using a common laptop computer, indicating great real-time ability with high performance.