Spaces:
Sleeping
Sleeping
| import numpy as np | |
| import supervision as sv | |
| from ultralytics import YOLO | |
| from tqdm import tqdm | |
| import re | |
| from collections import defaultdict | |
| from paddleocr import PaddleOCR | |
| from pdf2image import convert_from_path | |
| import json | |
| import cv2 | |
| import gradio as gr | |
| # Initialize YOLO model | |
| model_yolo = YOLO(model="yolov8n-box.pt") | |
| ocr = PaddleOCR(use_angle_cls=True, lang='en', use_gpu=False, show_log=False) | |
| def process_pdf(file): | |
| images = convert_from_path(file.name) | |
| # Function to process each slice of the image | |
| def slicer_callback(slice: np.ndarray) -> sv.Detections: | |
| result = model_yolo.predict(slice, conf=0.85)[0] | |
| detections = sv.Detections.from_ultralytics(result) | |
| return detections | |
| # Initialize the slicer | |
| slicer = sv.InferenceSlicer( | |
| callback=slicer_callback, | |
| slice_wh=(2000, 800), | |
| overlap_ratio_wh=(0.6, 0.6), | |
| overlap_filter_strategy=sv.OverlapFilter.NON_MAX_MERGE, | |
| iou_threshold=0.05, | |
| ) | |
| results = [] | |
| for pil_image in images: | |
| opencvImage = cv2.cvtColor(np.array(pil_image), cv2.COLOR_RGB2BGR) | |
| opencvImage = cv2.rotate(opencvImage, cv2.ROTATE_90_CLOCKWISE) | |
| # Perform inference on the entire image | |
| detections = slicer(opencvImage) | |
| # Function to run the TrOCR model with detections | |
| def run_example(detections): | |
| for detection in tqdm(detections): | |
| # Extract bounding box coordinates | |
| bbox = detection[0] | |
| x_min, y_min, x_max, y_max = bbox | |
| x_min, y_min, x_max, y_max = int(x_min), int(y_min), int(x_max), int(y_max) | |
| # Crop the detected region from the image | |
| cropped_image = opencvImage[y_min:y_max, x_min:x_max] | |
| result = ocr.ocr(cropped_image, cls=True)[0] | |
| if result is not None: | |
| text = '' | |
| if re.match(r"([A-Z])(\d+)-(\d+)", result[0][1][0]): | |
| text = result[0][1][0] | |
| elif re.match(r"([A-Z])(\d+)-(\d+)", ''.join([line[1][0] for line in result])): | |
| text = ''.join([line[1][0] for line in result]) | |
| # Print the generated text | |
| results.append(text) | |
| # Run example with detections | |
| run_example(detections) | |
| detected_numbers = defaultdict(list) | |
| for result in results: | |
| match = re.match(r"([A-Z])(\d+)-(\d+)", result) | |
| if match: | |
| letter = match.group(1) | |
| x = int(match.group(2)) | |
| y = int(match.group(3)) | |
| detected_numbers[(letter, x)].append(y) | |
| # Generate the desired JSON output | |
| output = {} | |
| for (letter, x) in sorted(detected_numbers.keys()): | |
| key = f"CB-{letter}{x}" | |
| value = [f"{letter}{x}-{i}" for i in sorted(detected_numbers[(letter, x)])] | |
| output[key] = value | |
| return json.dumps(output, indent=4) | |
| # Create the Gradio interface | |
| iface = gr.Interface( | |
| fn=process_pdf, | |
| inputs=gr.File(label="Upload PDF"), | |
| outputs="json", | |
| title="Extract Data from PDF", | |
| description="Upload a PDF file and get the JSON output of detected numbers." | |
| ) | |
| # Launch the Gradio app | |
| iface.launch() | |