Chinese text generation method with image signal feedback

The invention discloses a Chinese text generation method with image signal feedback, which comprises the following steps: word vector generation: generating binary images from all words in a dictionary required in a bert model, enabling each word to correspond to one binary image, enabling all images to pass through a CNN (Convolutional Neural Network), obtaining the last layer output by the network as a word vector, and generating a word vector; all the characters form an image information embedding matrix; sentence vector generation: during model training, an input text generates a binary image, the binary image passes through a CNN + RNN + CTC network, the input of RNN is taken as a sentence vector, which is equivalent to that word vectors are aggregated together, and then the word vectors are integrated into a real sentence vector. According to the method, CNN is adopted for extraction of single-character image signals, CNN + RNN + CTC optical character recognition models are adopted for extraction of who