How to use StyleText to generate text image with your own font？

Open goodmight opened this issue 2 years ago • 4 comments

请提供下述完整信息以便快速定位问题/Please provide the following information to quickly locate the problem

系统环境/System Environment：win10
版本号/Version：Paddle： PaddleOCR：问题相关组件/Related components：
运行指令/Command Code： python tools/synth_image.py -c configs/config.yml --style_image examples/style_images/2.jpg --text_corpus 123456 --language art_font
完整报错/Complete Error Message： No Errors but model predict（synth image） gives bad result with art font.

Here is what I've found.

In StyleText/engine/synthesisers.py, 40:

def synth_image(self, corpus, style_input, language="en"):
    corpus_list, text_input_list = self.text_drawer.draw_text(              # <---------- this draw_text work fine with font : **C:\Windows\fonts\complex_.ttf**
        corpus, language, style_input_width=style_input.shape[1])
    synth_result = self.predictor.predict(style_input, text_input_list)  # <----------   this predict which synth the text image with the background fails. It generates Korean text image with English or digits corpus.
    return synth_result

textImage generated in digits. fake_inprogress fake_inprogress2

but synth image is in Korean. fake_text fake_fusion

How to train the downloaded model with new fonts? Thanks! the download url is: https://paddleocr.bj.bcebos.com/dygraph_v2.0/style_text/style_text_models.zip as given in the StyleText/README.md