import os os.system('git clone https://github.com/facebookresearch/detectron2.git') os.system('pip install -e detectron2') os.system("git clone https://github.com/microsoft/unilm.git") os.system("sed -i 's/from collections import Iterable/from collections.abc import Iterable/' unilm/dit/object_detection/ditod/table_evaluation/data_structure.py") os.system("curl -LJ -o publaynet_dit-b_cascade.pth 'https://layoutlm.blob.core.windows.net/dit/dit-fts/publaynet_dit-b_cascade.pth?sv=2022-11-02&ss=b&srt=o&sp=r&se=2033-06-08T16:48:15Z&st=2023-06-08T08:48:15Z&spr=https&sig=a9VXrihTzbWyVfaIDlIT1Z0FoR1073VB0RLQUMuudD4%3D'") import sys sys.path.append("unilm") sys.path.append("detectron2") import cv2 import filetype from PIL import Image import numpy as np from io import BytesIO from pdf2image import convert_from_bytes, convert_from_path import re import requests from urllib.parse import urlparse, parse_qs from unilm.dit.object_detection.ditod import add_vit_config import torch from detectron2.config import CfgNode as CN from detectron2.config import get_cfg from detectron2.utils.visualizer import ColorMode, Visualizer from detectron2.data import MetadataCatalog from detectron2.engine import DefaultPredictor from huggingface_hub import hf_hub_download import gradio as gr # Step 1: instantiate config cfg = get_cfg() add_vit_config(cfg) #cfg.merge_from_file("cascade_dit_base.yml") cfg.merge_from_file("unilm/dit/object_detection/publaynet_configs/cascade/cascade_dit_base.yaml") # Step 2: add model weights URL to config filepath = hf_hub_download(repo_id="Sebas6k/DiT_weights", filename="publaynet_dit-b_cascade.pth", repo_type="model") cfg.MODEL.WEIGHTS = filepath # Step 3: set device cfg.MODEL.DEVICE = "cuda" if torch.cuda.is_available() else "cpu" # Step 4: define model predictor = DefaultPredictor(cfg) def analyze_image(img): md = MetadataCatalog.get(cfg.DATASETS.TEST[0]) if cfg.DATASETS.TEST[0]=='icdar2019_test': md.set(thing_classes=["table"]) else: md.set(thing_classes=["text","title","list","table","figure"]) ## these are categories from PubLayNet (PubMed PDF/XML data): https://ieeexplore.ieee.org/document/8977963 outputs = predictor(img) instances = outputs["instances"] # Ensure we're operating on CPU for numpy compatibility instances = instances.to("cpu") # Filter out figures based on class labels high_confidence = [] medium_confidence = [] low_confidence = [] for i in range(len(instances)): if md.thing_classes[instances.pred_classes[i]] == "figure": box = instances.pred_boxes.tensor[i].numpy().astype(int) cropped_img = img[box[1]:box[3], box[0]:box[2]] confidence_score = instances.scores[i].numpy() * 100 # convert to percentage confidence_text = f"Score: {confidence_score:.2f}%" # Overlay confidence score on the image # Enhanced label visualization with orange color font_scale = 0.9 font_thickness = 2 text_color = (255, 255, 255) # white background background_color = (255, 165, 0) # RGB for orange (text_width, text_height), _ = cv2.getTextSize(confidence_text, cv2.FONT_HERSHEY_SIMPLEX, font_scale, font_thickness) padding = 12 text_offset_x = padding - 3 text_offset_y = cropped_img.shape[0] - padding + 2 box_coords = ((text_offset_x, text_offset_y + padding // 2), (text_offset_x + text_width + padding, text_offset_y - text_height - padding // 2)) cv2.rectangle(cropped_img, box_coords[0], box_coords[1], background_color, cv2.FILLED) cv2.putText(cropped_img, confidence_text, (text_offset_x, text_offset_y), cv2.FONT_HERSHEY_SIMPLEX, font_scale, text_color, font_thickness) # Categorize images based on confidence levels if confidence_score > 85: high_confidence.append(cropped_img) elif confidence_score > 50: medium_confidence.append(cropped_img) else: low_confidence.append(cropped_img) v = Visualizer(img[:, :, ::-1], md, scale=1.0, instance_mode=ColorMode.SEGMENTATION) result_image = v.draw_instance_predictions(instances).get_image()[:, :, ::-1] return result_image, high_confidence, medium_confidence, low_confidence # output = predictor(img)["instances"] # v = Visualizer(img[:, :, ::-1], # md, # scale=1.0, # instance_mode=ColorMode.SEGMENTATION) # result = v.draw_instance_predictions(output.to("cpu")) # result_image = result.get_image()[:, :, ::-1] # ## figs = [img[box[1]:box[3], box[0]:box[2]] for box, cls in zip(output.pred_boxes, output.pred_classes) if md.thing_classes[cls] == "figure"] # # return result_image, figs def handle_input(input_data): images = [] #input_data is a dict with keys 'text' and 'files' if 'text' in input_data and input_data['text']: input_text = input_data['text'].strip() # this is either a URL or a PDF ID if input_text.startswith('http://') or input_text.startswith('https://'): # Extract the ID from the URL url_parts = urlparse(input_text) query_params = parse_qs(url_parts.fragment) # Assumes ID is a fragment parameter pdf_id = query_params.get('id', [None])[0] if not pdf_id: raise ValueError("PDF ID not found in URL") else: # Assume input is a direct PDF ID pdf_id = input_text if not re.match(r'^[a-zA-Z]{4}\d{4}$', pdf_id): raise ValueError("Invalid PDF ID format. Expected four letters followed by four numbers.") # Assume input is a PDF ID, convert to URL # Now construct the download URL pdf_url = construct_download_url(pdf_id) #https://download.industrydocuments.ucsf.edu/k/t/k/l/ktkl0236/ktkl0236.pdf # Assume input is a PDF URL pdf_data = download_pdf(pdf_url) images = pdf_to_images(pdf_data) if 'files' in input_data and input_data['files']: for file_path in input_data['files']: print("Type of file as uploaded:", type(file_path)) print(f" File: {file_path}") # Check if the input is a file and determine its type kind = filetype.guess(file_path) if kind.mime.startswith('image'): # Process a single image images.append(load_image(file_path)) # Process image directly elif kind.mime == 'application/pdf': # Convert PDF pages to images images.extend(pdf_to_images(file_path)) else: raise ValueError("Unsupported file type.") if not images: raise ValueError("No valid input provided. Please upload a file or enter a PDF ID.") # Assuming processing images returns galleries of images by confidence return process_images(images) def load_image(img_path): print(f"Loading image: {img_path}") # Load an image from a file path image = Image.open(img_path) if isinstance(image, Image.Image): image = np.array(image) # Convert PIL Image to numpy array # Ensure the image is in the correct format if image.ndim == 2: # Image is grayscale image = cv2.cvtColor(image, cv2.COLOR_GRAY2BGR) elif image.ndim == 3 and image.shape[2] == 3: image = cv2.cvtColor(image, cv2.COLOR_RGB2BGR) # image = image[:, :, ::-1] # Convert RGB to BGR if necessary return image def construct_download_url(pdf_id): # Construct the download URL from the PDF ID # https://download.examples.edu/k/t/k/l/ktkl0236/ktkl0236.pdf path_parts = '/'.join(pdf_id[i] for i in range(4)) # 'k/t/k/l' download_url = f"https://download.industrydocuments.ucsf.edu/{path_parts}/{pdf_id}/{pdf_id}.pdf" return download_url def download_pdf(pdf_url): # Download the PDF file from the given URL response = requests.get(pdf_url) response.raise_for_status() # Ensure we notice bad responses return BytesIO(response.content) def pdf_to_images(data_or_path): # Create a temporary directory to store the page images temp_dir = "temp_images" os.makedirs(temp_dir, exist_ok=True) try: # Convert PDF to a list of PIL images # Handle both BytesIO and file path input for PDF conversion if isinstance(data_or_path, BytesIO): # Convert directly from bytes pages = convert_from_bytes(data_or_path.read()) elif isinstance(data_or_path, str): # Convert from a file path pages = convert_from_path(data_or_path) # Save each page as an image file page_images = [] for i, page in enumerate(pages): image_path = os.path.join(temp_dir, f"page_{i+1}.jpg") page.save(image_path, "JPEG") page_images.append(load_image(image_path)) return page_images except Exception as e: print(f"Error converting PDF to images: {str(e)}") return [] finally: # Clean up the temporary directory (optional) # os.rmdir(temp_dir) pass def process_images(images): all_processed_images = [] all_high_confidence = [] all_medium_confidence = [] all_low_confidence = [] for img in images: #print("Type of img before processing:", type(img)) #print(f" img before processing: {img}") processed_images, high_confidence, medium_confidence, low_confidence = analyze_image(img) all_processed_images.append(processed_images) all_high_confidence.extend(high_confidence) all_medium_confidence.extend(medium_confidence) all_low_confidence.extend(low_confidence) return all_processed_images, all_high_confidence, all_medium_confidence, all_low_confidence title = "OIDA Image Collection Interactive demo: Document Layout Analysis with DiT and PubLayNet" description = "

OIDA Demo -- adapted liberally from https://huggingface.co/spaces/nielsr/dit-document-layout-analysis

Demo for Microsoft's DiT, the Document Image Transformer for state-of-the-art document understanding tasks. This particular model is fine-tuned on PubLayNet, a large dataset for document layout analysis (read more at the links below). To use it, simply upload an image or use the example image below and click 'Submit'. Results will show up in a few seconds. If you want to make the output bigger, right-click on it and select 'Open image in new tab'." article = "

Paper | Github Repo | HuggingFace doc | PubLayNet paper

" #examples =[['fpmj0236_Page_012.png'],['fnmf0234_Page_2.png'],['publaynet_example.jpeg'],['fpmj0236_Page_018.png'],['lrpw0232_Page_14.png'],['kllx0250'],['https://www.industrydocuments.ucsf.edu/opioids/docs/#id=yqgg0230']] examples =[{'files': ['fnmf0234_Page_2.png']},{'files': ['fpmj0236_Page_012.png']},{'files': ['lrpw0232.pdf']},{'text': 'https://www.industrydocuments.ucsf.edu/opioids/docs/#id=yqgg0230'},{'files':['fpmj0236_Page_018.png']},{'files':['lrpw0232_Page_14.png']},{'files':['publaynet_example.jpeg']},{'text':'kllx0250'},{'text':'txhk0255'}] #txhk0255 css = ".output-image, .input-image, .image-preview {height: 600px !important} td.textbox {display:none;} #component-5 .submit-button {display:none;}" #iface = gr.Interface(fn=handle_input, # inputs=gr.MultimodalTextbox(interactive=True, # label="Upload image/PDF file OR enter OIDA ID or URL", # file_types=["image",".pdf"], # placeholder="Upload image/PDF file OR enter OIDA ID or URL"), # outputs=[gr.Gallery(label="annotated documents"), # gr.Gallery(label="Figures with High (>85%) Confidence Scores"), # gr.Gallery(label="Figures with Moderate (50-85%) Confidence Scores"), # gr.Gallery(label="Figures with Lower Confidence (under 50%) Scores")], # title=title, # description=description, # examples=examples, # article=article, # css=css) ## enable_queue=True) with gr.Blocks(css=css) as iface: gr.Markdown(f"# {title}") gr.HTML(description) with gr.Row(): with gr.Column(): input = gr.MultimodalTextbox(interactive=True, label="Upload image/PDF file OR enter OIDA ID or URL", file_types=["image",".pdf"], placeholder="Upload image/PDF file OR enter OIDA ID or URL", submit_btn=None) submit_btn = gr.Button("Submit") gr.HTML('

') gr.Examples(examples, [input]) with gr.Column(): outputs = [gr.Gallery(label="annotated documents"), gr.Gallery(label="Figures with High (>85%) Confidence Scores"), gr.Gallery(label="Figures with Moderate (50-85%) Confidence Scores"), gr.Gallery(label="Figures with Lower Confidence (under 50%) Scores")] with gr.Row(): gr.HTML(article) submit_btn.click(handle_input, [input], outputs) iface.launch(debug=True, auth=[("oida", "OIDA3.1"), ("Brian", "Hi")]) #, cache_examples=True)