High Fidelity Scene Text Synthesis

View the PDF file for the paper entitled Dreamtext: Small text creation, by Yibin Wang, Weizhong Zhang, Honghui Xu and Cheng Jin

PDF HTML (experimental) view

a summary:The synthesis of the scene text includes providing specific texts on arbitrary images. The current methods usually formulate this task in a comprehensive way, but it lacks effective guidance at the level of personality during training. Moreover, its text, which was previously trained in one line type, is struggling to adapt to the varied lines patterns it faced in practical applications. Consequently, these methods suffer from distorting personality, repetition and absence, especially in multi -program scenarios. To this end, this paper suggests the text of Dreamtetex to synthesize the text of a high -resolution scene. Our main idea is to rebuild the process of spreading, introducing more directions designed by this task, to expose the attention of the model and correct it at the level of personality and enhance the learning of text areas. This transformation is an improved hybrid challenge, which includes both separate and continuous variables. To effectively address this challenge, we use an alternative improvement strategy. Meanwhile, we are jointly born and born to learn and use the diverse line in the training data set. This joint training was smoothly integrated into the alternative improvement process, which enhances a synergy relationship between the inclusion of the learning personality and the re -appreciation of the personality. Specifically, in each step, we first encrypt the potential position information that is created from letters from maps joining the form to latent letters. Then these masks are used to update the representation of specific letters in the current step, which in turn enables the generator to correct the personality of the personality in the subsequent steps. Each of the qualitative and quantitative results shows our way of the latest in art.