TextPage creation for other OCR engine

[darwinharianto]

I have tried tesseract-ocr-jpn to create textpage, but since the result isn't as good as I expected, is it possible to swap out to other engines?

I wanted to use easyOCR as my textpage back-end, but I am not sure how to construct the TextPage class itself.

What kind of data or structure do I need to prepare for constructing TextPage object?

[JorjMcKie]

This is not supported, sorry.
have you tried to increase the dpi value when using page.get_textpage_ocr()?

[darwinharianto]

Thanks you!
cjk.text_length(...) really helps with the character spacing!

I plan to show the OCR confidence using opacity. They are perfectly shown using inset_text(xxx_opacity= confidence), but how can I extract the opacity data?
Is it only available inside page.get_texttrace()?
can I get it from textpage?
I can't find opacity key inside textpage.extractRawDict

[JorjMcKie]

Is it only available inside page.get_texttrace()?

Yes, unfortunately this is one of the properties not available in TextPage, which is a MuPDF object and the data source for all page.get_text() and page.search_for() methods and variants. Although that may change in the future, there are no plans to do so to my best knowledge.

page.get_texttrace() is not based on TextPage, but directly walks through the page appearance instructions. Therefore, it can potentially return all properties of text like opacity (for both, fill and stroke operations) and text rendering mode. On the other hand it does lack the advanced text preprocessing features of TextPage, like the block-line-span-char hierarchy, dynamic space character generation and more.

[darwinharianto]

As a closing remarks...
to create a new textpage based on the ocr-ed page

def calculate_font_size(rect:Rect, text:str, font, rotation:float):
    # there are other faster methodd
    for i in range(1000):
        length = font.text_length(text, fontsize=i)
        if length > rect.width and (rotation == 0 or rotation == 180):
            return i-1
        elif length > rect.height and (rotation == 90 or rotation == 270):
            return i-1
    return I

newPage = doc.new_page(width=page.rect.width, height=page.rect.height)
page.set_rotation(page.rotation)
page.insert_font(fontname="cjk", fontbuffer=self.font.buffer)
shape = page.new_shape()
font = fitz.Font(fontname="cjk")

# this OCR results is from easyocr
for ocrResult in ocrResults:
      rect = fitz.Rect(ocrResult[0][0] + ocrResult[0][2])
      text = ocrResult[1]
      confidence = ocrResult[2]
      fontsize = OCRPDFProc.calculate_font_size(rect, text, font, rotation=0)
      # we can represent confidence with opacity
      shape.insert_text(rect.bl, text, fontsize, fontname='cjk', rotate=0, stroke_opacity=confidence, fill_opacity=confidence)
shape.commit()

[hiroi-sora]

Thank you very much!

I am the author of Umi-OCR, currently developing a PDF recognition feature to convert scanned documents into searchable PDFs. My project also uses other OCR engines, such as PaddleOCR-json.

Your idea has provided me with great inspiration; the positioning of the generated invisible text blocks is very accurate. Below is my improved calculate_font_size function, which can reduce the number of searches:

    # Calculate the font size that fills the width and height for a line of text
    def _calculateFontSize(self, text, w, h):
        if h > w:  # For vertical text, switch to horizontal for calculation
            w, h = h, w
        fontsize = round(h)  # Initial font size, assumed to be line height
        minSize = 5  # Minimum size limit
        getLen = lambda text, s: self.font.text_length(text, fontsize=s)
        while getLen(text, fontsize) < w:
            fontsize += 1  # Try increasing the font size until the line width just exceeds the limit
        fontsize -= 1
        while getLen(text, fontsize) > w and fontsize >= minSize:
            fontsize -= 1  # Try decreasing the font size until the line width is just less than the limit
        return fontsize

[nissansz]

Thanks you! cjk.text_length(...) really helps with the character spacing!

I plan to show the OCR confidence using opacity. They are perfectly shown using inset_text(xxx_opacity= confidence), but how can I extract the opacity data? Is it only available inside page.get_texttrace()? can I get it from textpage? I can't find opacity key inside textpage.extractRawDict

实际显示的效果是怎么样的？发个图片看看