from typing import Tuple, List, Sequence, Optional, Union
from torchvision import transforms
from torch import nn, Tensor
from PIL import Image
from pathlib import Path
from bs4 import BeautifulSoup as bs
import numpy as np
import numpy.typing as npt
from numpy import uint8
ImageType = npt.NDArray[uint8]
from transformers import AutoModelForObjectDetection
import torch
import matplotlib.pyplot as plt
import matplotlib.patches as patches
from matplotlib.patches import Patch
from utils import draw_only_box
from unitable import UnitablePredictor
from ultralyticsplus import YOLO, render_result
from doctrfiles import DoctrWordDetector,DoctrTextRecognizer
from utils import crop_an_Image,cropImageExtraMargin
from utils import denoisingAndSharpening
"""
USES YOLO FOR DETECITON INSTEAD OF TABLE TRANSFORMER
Table TransFORMER
"""
html_table_template = (
lambda table: f"""
"""
)
class DetectionAndOcrTable3():
#This components can take in entire pdf page as input , scan for tables and return the table in html format
#Uses the full unitable model - different to DetectionAndOcrTable1
def __init__(self,englishFlag = True):
self.unitablePredictor = UnitablePredictor()
self.detector = YOLO('foduucom/table-detection-and-extraction')
# set model parameters
self.detector.overrides['conf'] = 0.25 # NMS confidence threshold
self.detector.overrides['iou'] = 0.45 # NMS IoU threshold
self.detector.overrides['agnostic_nms'] = False # NMS class-agnostic
self.detector.overrides['max_det'] = 1000 # maximum number of detections per image
self.wordDetector = DoctrWordDetector(architecture="db_resnet50",
path_weights="doctrfiles/models/db_resnet50-79bd7d70.pt",
path_config_json ="doctrfiles/models/db_resnet50_config.json")
if englishFlag:
self.textRecognizer = DoctrTextRecognizer(architecture="master", path_weights="./doctrfiles/models/master-fde31e4a.pt",
path_config_json="./doctrfiles/models/master.json")
else:
self.textRecognizer = DoctrTextRecognizer(architecture="parseq", path_weights="./doctrfiles/models/doctr-multilingual-parseq.bin",
path_config_json="./doctrfiles/models/multilingual-parseq-config.json")
@staticmethod
def save_detection(detected_lines_images:List[ImageType], prefix = './res/test1/res_'):
i = 0
for img in detected_lines_images:
pilimg = Image.fromarray(img)
pilimg.save(prefix+str(i)+'.png')
i=i+1
@staticmethod
def build_table_from_html_and_cell(
structure: List[str], content: List[str] = None
) -> List[str]:
"""Build table from html and cell token list"""
assert structure is not None
html_code = list()
# deal with empty table
if content is None:
content = ["placeholder"] * len(structure)
for tag in structure:
if tag in ("| [] | ", ">[]"):
if len(content) == 0:
continue
cell = content.pop(0)
html_code.append(tag.replace("[]", cell))
else:
html_code.append(tag)
return html_code
"""
Valid 'Boxes' object attributes and properties are:
Attributes:
boxes (torch.Tensor) or (numpy.ndarray): A tensor or numpy array containing the detection boxes,
with shape (num_boxes, 6).
orig_shape (torch.Tensor) or (numpy.ndarray): Original image size, in the format (height, width).
Properties:
xyxy (torch.Tensor) or (numpy.ndarray): The boxes in xyxy format.
conf (torch.Tensor) or (numpy.ndarray): The confidence values of the boxes.
cls (torch.Tensor) or (numpy.ndarray): The class values of the boxes.
xywh (torch.Tensor) or (numpy.ndarray): The boxes in xywh format.
xyxyn (torch.Tensor) or (numpy.ndarray): The boxes in xyxy format normalized by original image size.
xywhn (torch.Tensor) or (numpy.ndarray): The boxes in xywh format normalized by original image size.
"""
# Image is page image
def predict(self,image:Image.Image,debugfolder_filename_page_name = None,denoise =False):
results = self.detector.predict(image)
#Array of bboxes
bbxs = results[0].boxes.xyxy.int().tolist()
#Array of confidences
conf = results[0].boxes.conf.float().tolist()
print(bbxs)
print(conf)
#images_to_recognizer = cropImage(bxs, img)
img_to_save = draw_only_box(image, bbxs)
img_to_save.save(debugfolder_filename_page_name+"detectionBoxRes.png", quality=95)
# we need something to draw the detection
cropped_tables =[]
for i in range (len(bbxs)):
# TODO: find the right confidence and padding values
if conf[i]< 0.65:
continue
padded = [bbxs[i][0]-10,bbxs[i][1]-10,bbxs[i][2]+10,bbxs[i][3]+10]
cropped_table = image.convert("RGB").crop(padded)
cropped_table.save(debugfolder_filename_page_name +"yolo_cropped_table_"+str(i)+".png")
cropped_tables.append(cropped_table)
print("number of cropped tables found: "+str(len(cropped_tables)))
# Step 1: Unitable
#This take PIL Images as input
if cropped_tables != []:
if denoise:
cropped_tables =denoisingAndSharpening(cropped_tables)
pred_htmls, pred_bboxs = self.unitablePredictor.predict(cropped_tables,debugfolder_filename_page_name)
table_codes = []
for k in range(len(cropped_tables)):
pred_html =pred_htmls[k]
pred_bbox = pred_bboxs[k]
# Some tabless have a lot of words in their header
# So for the headers, give doctr word ddetector doesn't work when the images aren't square
table_header_cells = 0
header_exists = False
for cell in pred_html:
if cell=='>[]' or cell == '[] | ':
table_header_cells += 1
if cell =='':
header_exists = True
break
if not header_exists:
table_header_cells = 0
pred_cell = []
cell_imgs_to_viz = []
cell_img_num=0
# Find what one line should be if there is a cell with a single line
one_line_height = 100000
for i in range(table_header_cells):
box = pred_bbox[i]
xmin, ymin, xmax, ymax = box
current_box_height = abs(ymax-ymin)
if current_box_height 0 and current_box_height>one_line_height+5:
cell_img= cropImageExtraMargin([fourbytwo],cropped_tables[k],margin=1.4)[0]
table_header_cells -= 1
#List of 4 x 2
detection_results = self.wordDetector.predict(cell_img,sort_vertical=True)
input_to_recog = []
if detection_results == []:
input_to_recog.append(cell_img)
else:
for wordbox in detection_results:
cropped_image= crop_an_Image(wordbox.box,cell_img)
if cropped_image.shape[0] >0 and cropped_image.shape[1]>0:
input_to_recog.append(cropped_image)
else:
print("Empty image")
else:
cell_img = crop_an_Image(fourbytwo,cropped_tables[k])
if table_header_cells>0:
table_header_cells -= 1
if cell_img.shape[0] >0 and cell_img.shape[1]>0:
input_to_recog =[cell_img]
cell_imgs_to_viz.append(cell_img)
if input_to_recog != []:
words = self.textRecognizer.predict_for_tables(input_to_recog)
cell_output = " ".join(words)
pred_cell.append(cell_output)
else:
#Don't lose empty cell
pred_cell.append("")
#self.save_detection(cell_imgs_to_viz,prefix = './res/test4/cell_imgs_')
print(pred_cell)
#Step3 :
pred_code = self.build_table_from_html_and_cell(pred_html, pred_cell)
pred_code = "".join(pred_code)
pred_code = html_table_template(pred_code)
soup = bs(pred_code)
#formatted and indented) string representation of the HTML document
table_code = soup.prettify()
print(table_code)
table_codes.append(table_code)
return table_codes
return []