Abstract: Aiming at the problem of low recognition rate and instability caused by the interference of complex backgrounds, ambient light and other factors in the current bank card number recognition, a bank card number recognition method based on YOLOv3 and MobileNetv2 is proposed. Firstly, we preprocess the datasets, collect multi-style bank card images, batch Laplace sharpening and partial image enhancement, and label the images. Afterwards we need to construct the YOLOv3 region segmentation network structure, input the marked data set into the YOLOv3 network, optimize the target size loss and focal loss, optimize the confidence loss, control the iterative calculation, segment the word code region, output the preliminary recognition model and calculate the preliminary recognition results. Then, the improved YOLOv3 network and basic MobileNetv2 network are constructed, the preprocessed data set is input for training, the joint recognition model is output, and the joint recognition results are calculated. Finally, by comparing the preliminary recognition results with the joint recognition results, the results with the highest accuracy are output. During the experiment, the data set of China Software Cup and Heywhale bank card are integrated and expanded, and divided into four types according to the word code form, and the positioning identification and accuracy comparison experiment are conducted respectively. Experimental results show that in terms of the localization effect of the code area, the proposed method is superior to the traditional CNN network and the basic YOLOv3 network. In terms of the accuracy of the code recognition, the accuracy of the method proposed on four types of bank cards is 93.74%, 93.21%, 95.14% and 99.10% respectively, all of which perform better than the improved YOLOv3 and YOLOv3-MobileNetv2 methods. Experiments prove that proposed method can recognize bank card codes with different code patterns under complex background, overcome the influence of environmental factors on card number recognition, improve recognition accuracy, and have good robustness. And it runs stably and reliably on the designed verification system and application platform.