Method for generating picture by characters based on bidirectional architecture generative adversarial network

Abstract

The invention discloses a method for generating a picture by characters based on a bidirectional architecture generative adversarial network. The method comprises the steps: firstly, employing a pre-trained character coding network for analyzing text meanings, and mapping the text meanings to a semantic vector space; and then enabling a bidirectional architecture adversarial generative network model to generate a picture corresponding to the semantic vector by utilizing the semantic vector. Comparing with the related art, the method has the following advantages that the method employs the bidirectional architecture ideas, and achieves a process of generating a high-quality picture only by depending on a text in combination with an adversarial generation network; and the purpose of enhancing the semantic consistency between the picture and the text is achieved by improving the attention mechanism and adjusting the batch numeration, and experiments prove that the model architecture can significantly improve the quality and diversity of the synthesized picture.

Description

【Technical field】 [0001] The invention relates to the technical field of word processing, in particular to a method for generating pictures from words based on a two-way architecture confrontation generation network. 【Background technique】 [0002] Generating a picture corresponding to it and in line with reality based on the semantics of the text involves many fields. The technology in the field of natural language processing is required in the word processing step, and the relevant knowledge of computer vision is required when generating pictures. Therefore, most of the existing technologies are divided into two parts to achieve: [0003] 1. Text encoding, this step generally uses two neural networks to process text and pictures separately, and maps them to the same vector space through learning. For example, in the AttnGAN network, a CNN network is used to process pictures, an LSTM network is used to process text, and then a specific objective function is used to optimiz...

AI Technical Summary

Problems solved by technology

However, there is still an insurmountable gap between the semantics of the text itself and the visual content of the picture itself, making it difficult to establish semantic consistency

Qiao et al. solved this problem by introducing an image annotation model, but this method was too dependent on the performance of the image annotation model; Yin et al. used siamese networks to directly extract semantic consistency from text descriptions, but this method was too Rely on the text richness of the training dataset, which is not ideal in actual use

[0006]In general, existing methods have certain limitations and are difficult to promote

Images