Research on Recognition of Official Script in Natural Environment

The domestic dual collaborative education system is becoming increasingly mature, and the computer industry is a hot spot. Aiming at problems such as the absence of a dataset and the complicated recognition of pictures containing multiple characters, this paper mainly studies the preprocessing, data augmentation, word positioning and cutting methods of text pictures in a natural environment. First, image processing and data augmentation functions are used to create datasets; then, a CNN convolutional neural network model is built for training; finally, the method combining horizontal projection and vertical projection is used to locate and cut the word of multi-word graphs. The training accuracy was 96% and the test accuracy was 94%. The results show that the preprocessed image is in line with expectations, and the effect of the cutting monogram has a certain applicability.