A High-Precision Dot Matrix Character Recognition Method in Complex Background

Dot matrix character recognition is a challenging task with many difficulties, such as lack of datasets, complex background pattern, character distortion caused by uneven outer packaging of products. In this paper, we propose a new method of synthesizing dot matrix character image, which can automatically generate a large number of datasets and annotation files. The synthesized dot matrix character image is very similar to the real image. On this basis, we propose a high-precision dot matrix character recognition method in complex background, which constructs a new neural network architecture by inserting the spatial transformer network (STN) into the convolutional recurrent neural network (CRNN). The STN can transform an input image to a rectified image by performing explicit spatial transformations, which can help to improve the performance of dot matrix character recognition. Experiments show that the proposed method achieves the state-of-the-art recognition result, without the loss of the inference speed.