app.py

import numpy as np
import PIL
from PIL import Image, ImageDraw, ImageFont
import gradio as gr
import torch
import easyocr
import os
from pathlib import Path
import cv2
import pandas as pd
from transformers import TrOCRProcessor, VisionEncoderDecoderModel


#torch.hub.download_url_to_file('https://github.com/AaronCWacker/Yggdrasil/blob/main/images/BeautyIsTruthTruthisBeauty.JPG', 'BeautyIsTruthTruthisBeauty.JPG')
#torch.hub.download_url_to_file('https://github.com/AaronCWacker/Yggdrasil/blob/main/images/PleaseRepeatLouder.jpg', 'PleaseRepeatLouder.jpg')
#torch.hub.download_url_to_file('https://github.com/AaronCWacker/Yggdrasil/blob/main/images/ProhibitedInWhiteHouse.JPG', 'ProhibitedInWhiteHouse.JPG')

torch.hub.download_url_to_file('https://raw.githubusercontent.com/AaronCWacker/Yggdrasil/master/images/20-Books.jpg','20-Books.jpg')
torch.hub.download_url_to_file('https://github.com/JaidedAI/EasyOCR/raw/master/examples/english.png', 'COVID.png')
torch.hub.download_url_to_file('https://github.com/JaidedAI/EasyOCR/raw/master/examples/chinese.jpg', 'chinese.jpg')
torch.hub.download_url_to_file('https://github.com/JaidedAI/EasyOCR/raw/master/examples/japanese.jpg', 'japanese.jpg')
torch.hub.download_url_to_file('https://i.imgur.com/mwQFd7G.jpeg', 'Hindi.jpeg')


def draw_boxes(image, bounds, color='yellow', width=2):
    draw = ImageDraw.Draw(image)
    for bound in bounds:
        p0, p1, p2, p3 = bound[0]
        draw.line([*p0, *p1, *p2, *p3, *p0], fill=color, width=width)
    return image

def box_size(box):
    points = box[0]
    if len(points) == 4:
        x1, y1 = points[0]
        x2, y2 = points[2]
        return abs(x1 - x2) * abs(y1 - y2)
    else:
        return 0

def box_position(box):
    return (box[0][0][0] + box[0][2][0]) / 2, (box[0][0][1] + box[0][2][1]) / 2

def filter_temporal_profiles(temporal_profiles, period_index):
    filtered_profiles = []
    for profile in temporal_profiles:
        filtered_profile = []
        for t, text in profile:
            # Remove all non-digit characters from text
            filtered_text = ''.join(filter(str.isdigit, text))
            # Insert period at the specified index
            filtered_text = filtered_text[:period_index] + "." + filtered_text[period_index:]
            try:
                filtered_value = float(filtered_text)
            except ValueError:
                continue
            filtered_profile.append((t, filtered_value))
        filtered_profiles.append(filtered_profile)
    return filtered_profiles


device = 'cuda' if torch.cuda.is_available() else 'cpu'
processor = TrOCRProcessor.from_pretrained('microsoft/trocr-large-printed')
model = VisionEncoderDecoderModel.from_pretrained('microsoft/trocr-large-printed').to(device)

def inference(video, lang, time_step, full_scan, number_filter, use_trocr, period_index):
    output = 'results.mp4'
    reader = easyocr.Reader(lang)
    bounds = []   
    vidcap = cv2.VideoCapture(video)
    success, frame = vidcap.read()
    count = 0
    frame_rate = vidcap.get(cv2.CAP_PROP_FPS)
    output_frames = []
    temporal_profiles = []
    compress_mp4 = False

    # Get the positions of the largest boxes in the first frame
    bounds = reader.readtext(frame)
    for i in reversed(range(len(bounds))):
        box = bounds[i]
        # Remove box if it doesn't contain a number
        if not any(char.isdigit() for char in box[1]):
            bounds.pop(i)
    im = PIL.Image.fromarray(frame)
    im_with_boxes = draw_boxes(im, bounds)
    largest_boxes = sorted(bounds, key=lambda x: box_size(x), reverse=True)
    positions = [box_position(b) for b in largest_boxes]
    temporal_profiles = [[] for _ in range(len(largest_boxes))]

    # Match bboxes to position and store the text read by OCR
    while success:
        if count % (int(frame_rate * time_step)) == 0:
            if full_scan:
                bounds = reader.readtext(frame)
                for box in bounds:
                    bbox_pos = box_position(box)
                    for i, position in enumerate(positions):
                        distance = np.linalg.norm(np.array(bbox_pos) - np.array(position))
                        if distance < 50:
                            temporal_profiles[i].append((count / frame_rate, box[1]))
                            break
            else:
                for i, box in enumerate(largest_boxes):
                    x1, y1 = box[0][0]
                    x2, y2 = box[0][2]
                    box_width = x2 - x1
                    box_height = y2 - y1
                    ratio = 0.2
                    x1 = max(0, int(x1 - ratio * box_width))
                    x2 = min(frame.shape[1], int(x2 + ratio * box_width))
                    y1 = max(0, int(y1 - ratio * box_height))
                    y2 = min(frame.shape[0], int(y2 + ratio * box_height))
                    cropped_frame = frame[y1:y2, x1:x2]
                    if use_trocr:
                        pixel_values = processor(images=cropped_frame, return_tensors="pt").pixel_values
                        generated_ids = model.generate(pixel_values.to(device))
                        generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
                        temporal_profiles[i].append((count / frame_rate, generated_text))
                    else:
                        text = reader.readtext(cropped_frame)
                        if text:
                            temporal_profiles[i].append((count / frame_rate, text[0][1]))
            
            im = PIL.Image.fromarray(frame)
            im_with_boxes = draw_boxes(im, bounds)
            output_frames.append(np.array(im_with_boxes))

        success, frame = vidcap.read()
        count += 1

    if number_filter:
        # Filter the temporal profiles by removing non-matching characters and converting to floats
        temporal_profiles = filter_temporal_profiles(temporal_profiles, int(period_index))

    # Default resolutions of the frame are obtained. The default resolutions are system dependent.
    # We convert the resolutions from float to integer.
    width = int(vidcap.get(cv2.CAP_PROP_FRAME_WIDTH))
    height = int(vidcap.get(cv2.CAP_PROP_FRAME_HEIGHT))
    fps = vidcap.get(cv2.CAP_PROP_FPS)
    frames_total = int(vidcap.get(cv2.CAP_PROP_FRAME_COUNT))

    # Define the codec and create VideoWriter object.
    if compress_mp4:
        temp = f"{Path(output).stem}_temp{Path(output).suffix}"
        output_video = cv2.VideoWriter(
            temp, cv2.VideoWriter_fourcc(*"mp4v"), fps, (width, height)
        )
    else:
        output_video = cv2.VideoWriter(output, cv2.VideoWriter_fourcc(*"mp4v"), fps, (width, height))
    for frame in output_frames:
        output_video.write(frame)

    # Draw boxes with box indices in the first frame of the output video
    im = Image.fromarray(output_frames[0])
    draw = ImageDraw.Draw(im)
    font_size = 30
    font_path = "/usr/share/fonts/truetype/dejavu/DejaVuSans.ttf"
    for i, box in enumerate(largest_boxes):
        draw.text((box_position(box)), f"Box {i+1}", fill='red', font=ImageFont.truetype(font_path, font_size))
    
    output_video.release()
    vidcap.release()

    if compress_mp4:
        # Compressing the video for smaller size and web compatibility.
        os.system(
            f"ffmpeg -y -i {temp} -c:v libx264 -b:v 5000k -minrate 1000k -maxrate 8000k -pass 1 -c:a aac -f mp4 /dev/null && ffmpeg -y -i {temp} -c:v libx264 -b:v 5000k -minrate 1000k -maxrate 8000k -pass 2 -c:a aac -movflags faststart {output}"
        )
        os.system(f"rm -rf {temp} ffmpeg2pass-0.log ffmpeg2pass-0.log.mbtree")
    
    # Format temporal profiles as a DataFrame
    df_list = []
    for i, profile in enumerate(temporal_profiles):
        for t, text in profile:
            df_list.append({"Box": f"Box {i+1}", "Time (s)": t, "Text": text})
        df_list.append({"Box": f"", "Time (s)": "", "Text": ""})
    df = pd.concat([pd.DataFrame(df_list)])
    return output, im, df


title = '🖼️Video to Multilingual OCR👁️Gradio'
description = 'Multilingual OCR which works conveniently on all devices in multiple languages. Adjust time-step for inference and the scan mode according to your requirement. For `Full Screen Scan`, model scan the whole image if flag is ture, while scan only the box detected at the first video frame; this accelerate the inference while detecting the fixed box.'
article = "<p style='text-align: center'></p>"

examples = [
['test.mp4',['en'],10,False,True,True,1]
]

css = ".output_image, .input_image {height: 40rem !important; width: 100% !important;}"
choices = [
    "ch_sim",
    "ch_tra",
    "de",
    "en",
    "es",
    "ja",
    "hi",
    "ru"
]


gr.Interface(
    inference,
    [
        gr.inputs.Video(label='Input Video'),
        gr.inputs.CheckboxGroup(choices, type="value", default=['en'], label='Language'),
        gr.inputs.Number(label='Time Step (in seconds)', default=1.0),
        gr.inputs.Checkbox(label='Full Screen Scan'),
        gr.inputs.Checkbox(label='Use TrOCR large (this is only available when Full Screen Scan is disable)'),
        gr.inputs.Checkbox(label='Number Filter (remove non-digit char and insert period)'),
        gr.inputs.Textbox(label="period position",default=1)
    ],
    [
        gr.outputs.Video(label='Output Video'),
        gr.outputs.Image(label='Output Preview', type='numpy'),
        gr.outputs.Dataframe(headers=['Box', 'Time (s)', 'Text'], type='pandas'),
    ],
    title=title,
    description=description,
    article=article,
    examples=examples,
    css=css,
    enable_queue=True
).launch(debug=True)