An open-source document parsing model with SOTA accuracy and speed, using an RSI loop during post-training.
import cv2
import numpy as np
def resize_image(image: np.ndarray, scale: float) -> np.ndarray:
"""Resize an image isotropically while keeping both dimensions nonzero."""
if scale <= 0:
raise ValueError("Image scale must be greater than zero.")
height, width = image.shape[:2]
target_size = (
max(1, round(width * scale)),
max(1, round(height * scale)),
)
return cv2.resize(image, target_size, interpolation=cv2.INTER_AREA)
def crop_image(image: np.ndarray, bbox: list[float]) -> np.ndarray:
"""
Crop an image using the given bounding box.
bbox: [x1, y1, x2, y2]
"""
# D-FINE emits floating-point pixel coordinates. Quantize them to the
# nearest pixel so recognition crops match the reference pipeline.
x1, y1, x2, y2 = (round(value) for value in bbox)
# Ensure coordinates are within image boundaries
h, w = image.shape[:2]
x1 = max(0, x1)
y1 = max(0, y1)
x2 = min(w, x2)
y2 = min(h, y2)
return image[y1:y2, x1:x2].copy()