app.py

import os
import shutil

import easyocr
import gradio as gr
import py3langid as langid
from PIL import Image
from transformers import M2M100ForConditionalGeneration, M2M100Tokenizer

languages = ['en', 'fr', 'es']
threshold = 0.2
langid.set_languages(languages)

def read_text_from_image(img):
  reader = easyocr.Reader(languages)
  result = reader.readtext(img)
  texts = []

  for (bbox, text, prob) in result:
    # print(f"Text: {text}, Probability: {prob}")
    # filter by prob
    if prob > threshold:
      texts.append(text)

  if len(texts) == 0:
    raise ValueError("No text detected")

  concatenated_text = " ".join(texts).lower()
  return concatenated_text

def detect_language(text):
  lang, prob = langid.classify(text)
  # print(f"The text {text} is classify as {lang} with probability {prob}")
  return lang

def translate_to_id(text, lang):
  query = f"translate from {lang} to indonesia: {text}"

  tokenizer = M2M100Tokenizer.from_pretrained("facebook/m2m100_1.2B", src_lang=lang)
  model = M2M100ForConditionalGeneration.from_pretrained("facebook/m2m100_1.2B")

  input_ids = tokenizer(query, return_tensors="pt").input_ids
  outputs = model.generate(input_ids, forced_bos_token_id=tokenizer.get_lang_id("id"))

  translated = tokenizer.decode(outputs[0], skip_special_tokens=True)
  return translated

def predict(img):
  try:
    text = read_text_from_image(img)
    lang = detect_language(text)
    translated = translate_to_id(text, lang)

    # print(f"Text: {text}. Language: {lang}. Translated: {translated}")
    return translated
  except ValueError as e:
    print(e)
    return e

app = gr.Interface(
    fn=predict,
    inputs=gr.Image(label="Input Image"),
    outputs='text',
)

app.launch()