File size: 2,739 Bytes
aa8cd87
0bbf6ef
 
e91a768
0380162
57d20a1
f496449
07f5bd9
 
aa8cd87
0bbf6ef
07f5bd9
43d306c
e91a768
 
 
 
aa8cd87
43d306c
879dfbb
c65777e
 
 
 
 
e91a768
 
 
 
c65777e
 
 
 
e91a768
b296597
3cadd69
4504622
3cadd69
 
 
 
 
 
 
c65777e
3cadd69
 
ec2e6e8
07f5bd9
 
 
ff2c42f
07f5bd9
ff2c42f
07f5bd9
ff2c42f
 
 
649e38b
ff2c42f
07f5bd9
ff2c42f
 
 
 
649e38b
89415f2
e91a768
0bbf6ef
 
 
e91a768
 
 
e0a154b
e91a768
0bbf6ef
 
 
 
03c8fc6
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82

import gradio as gr
import numpy as np
import fitz  # PyMuPDF
import spaces
from ultralytics import YOLOv10

# Load the trained model

model = YOLOv10("best.pt")


# Define the class indices for figures and tables
figure_class_index = 3  # class index for figures
table_class_index = 4   # class index for tables

# Function to perform inference on an image and return bounding boxes for figures and tables
@spaces.GPU
def infer_image_and_get_boxes(image, confidence_threshold=0.6):
    results = model.predict(image)
    boxes = [
        (int(box.xyxy[0][0]), int(box.xyxy[0][1]), int(box.xyxy[0][2]), int(box.xyxy[0][3]))
        for result in results for box in result.boxes
        if int(box.cls[0]) in {figure_class_index, table_class_index} and box.conf[0] > confidence_threshold
    ]
    return boxes

# Function to crop images from the boxes
def crop_images_from_boxes(image, boxes, scale_factor):
    cropped_images = [
        image[int(y1 * scale_factor):int(y2 * scale_factor), int(x1 * scale_factor):int(x2 * scale_factor)]
        for (x1, y1, x2, y2) in boxes
    ]
    return cropped_images

@spaces.GPU
def process_pdf(pdf_file):
    # Open the PDF file
    doc = fitz.open(pdf_file)
    all_cropped_images = []

    # Set the DPI for inference and high resolution for cropping
    low_dpi = 50
    high_dpi = 300

    # Calculate the scaling factor
    scale_factor = high_dpi / low_dpi

    # Pre-cache all page pixmaps at low DPI
    low_res_pixmaps = [page.get_pixmap(dpi=low_dpi) for page in doc]
    
    # Loop through each page
    for page_num, low_res_pix in enumerate(low_res_pixmaps):
        low_res_img = np.frombuffer(low_res_pix.samples, dtype=np.uint8).reshape(low_res_pix.height, low_res_pix.width, 3)
        
        # Get bounding boxes from low DPI image
        boxes = infer_image_and_get_boxes(low_res_img)
        
        if boxes:
            # Load high DPI image for cropping only if boxes are found
            high_res_pix = doc[page_num].get_pixmap(dpi=high_dpi)
            high_res_img = np.frombuffer(high_res_pix.samples, dtype=np.uint8).reshape(high_res_pix.height, high_res_pix.width, 3)
            
            # Crop images at high DPI
            cropped_imgs = crop_images_from_boxes(high_res_img, boxes, scale_factor)
            all_cropped_images.extend(cropped_imgs)
    
    return all_cropped_images

# Create Gradio interface
iface = gr.Interface(
    fn=process_pdf,
    inputs=gr.File(label="Upload a PDF"),
    outputs=gr.Gallery(label="Cropped Figures and Tables from PDF Pages"),
    title="Fast document layout analysis based on YOLOv10",
    description="Upload a PDF file to get cropped figures and tables from each page."
)

# Launch the app
iface.launch()