akaocr 2.1.6

akaOCR Package Tools

akaOCR

✨ Description

This package is compatible with akaOCR for ocr pipeline program (Text Detection, Text Recognition & Text Rotation), using ONNX format model (CPU & GPU speed can be x2 Times Faster). This code is referenced from this awesome repo.

🚀 Features

1. Text Detection.

from akaocr import BoxEngine
import cv2

# Load image
img_path = "path/to/image.jpg"
image = cv2.imread(img_path)

# Initialize text detector
box_engine = BoxEngine(
    model_path=None,             # Path to detection model
    side_len=None,               # Minimum image size for inference
    conf_thres=0.5,              # Confidence threshold
    mask_thes=0.4,               # Binarization threshold
    unclip_ratio=2.0,            # Margin for expanding box
    max_candidates=1000,         # Maximum number of boxes
    device='cpu'                 # 'cpu' or 'gpu'
)

# Run inference
results = box_engine(image)
# Output: List of bounding boxes as np.array([[x1, y1], [x2, y2], [x3, y3], [x4, y4]], dtype=np.float32)

2. Text Recognition.

from akaocr import TextEngine
import cv2

# Load cropped image
img_path = "path/to/cropped_image.jpg"
cropped_image = cv2.imread(img_path)

# Initialize text recognizer
text_engine = TextEngine(
    model_path=None,             # Path to recognition model
    vocab_path=None,             # Path to vocabulary file
    use_space_char=True,         # Include space in predictions
    batch_sizes=32,              # Batch size for inference
    model_shape=[3, 48, 320],    # Expected input shape [C, H, W]
    max_wh_ratio=None,           # Max width-height ratio for resizing
    device='cpu'                 # 'cpu' or 'gpu'
)

# Run inference
results = text_engine(cropped_image)
# Output: List of tuples: (recognized_text, confidence_score)

3. Text Rotation.

from akaocr import ClsEngine
import cv2

# Load cropped image
img_path = "path/to/cropped_image.jpg"
cropped_image = cv2.imread(img_path)

# Initialize orientation classifier
rotate_engine = ClsEngine(
    model_path=None,             # Path to rotation classification model
    conf_thres=0.75,             # Confidence threshold
    device='cpu'                 # 'cpu' or 'gpu'
)

# Run inference
results = rotate_engine(cropped_image)
# Output: List of tuples: (label: 0 or 180, confidence_score)

🔥 Usage

import numpy as np
import cv2

from akaocr import BoxEngine, TextEngine
from typing import List, Tuple


class Pipeline:
    def __init__(self, device: str = 'cpu'):
        # Initializes the OCR pipeline. The computation device to use ('cpu' or 'gpu').

        self.box_engine = BoxEngine(device=device)
        self.text_engine = TextEngine(device=device)

    @staticmethod
    def _transform_image(image: np.ndarray, box: np.ndarray) -> np.ndarray:
        """
        Applies a perspective transform to straighten a detected text box.

        Args:
            image (np.ndarray): The source image.
            box (np.ndarray): A 4x2 numpy array of corner points for the text box.

        Returns:
            np.ndarray: The cropped and straightened image of the text region.
        """
        if not isinstance(box, np.ndarray) or box.shape != (4, 2):
            raise ValueError("Input 'box' must be a 4x2 NumPy array.")

        # Ensure points are float32 for cv2 functions
        box = box.astype(np.float32)

        # Calculate the width and height of the destination image
        width = int(max(np.linalg.norm(box[0] - box[1]), np.linalg.norm(box[2] - box[3])))
        height = int(max(np.linalg.norm(box[0] - box[3]), np.linalg.norm(box[1] - box[2])))

        # Define the destination points for a standard rectangle
        dst_pts = np.array([
            [0, 0],
            [width - 1, 0],
            [width - 1, height - 1],
            [0, height - 1]
        ], dtype=np.float32)

        # Get the perspective transform matrix and apply it
        matrix = cv2.getPerspectiveTransform(box, dst_pts)
        warped_img = cv2.warpPerspective(
            image, matrix, (width, height),
            borderMode=cv2.BORDER_REPLICATE,
            flags=cv2.INTER_CUBIC
        )
        return warped_img

    def __call__(self, image: np.ndarray) -> List[Tuple[np.ndarray, str]]:
        """
        Processes an image to detect and recognize text.

        Args:
            image (np.ndarray): The input image in BGR format.

        Returns:
            List[Tuple[np.ndarray, str]]: A list of tuples, where each tuple
            contains the bounding box (4x2 array) and the recognized text.
        """
        print("Starting text detection...")
        boxes = self.box_engine(image)
        if not boxes:
            print("No text boxes detected.")
            return []
        
        print(f"Detected {len(boxes)} text boxes.")

        # Prepare all cropped images for batch processing
        transformed_images = [self._transform_image(image, box) for box in boxes]

        print("Starting text recognition on detected boxes...")
        texts = self.text_engine(transformed_images)
        print("Text recognition complete.")

        # Combine boxes with their corresponding recognized text
        return list(zip(boxes, texts))

Note: akaOCR (Transform documents into useful data with AI-based IDP - Intelligent Document Processing) - helps make inefficient manual entry a thing of the past—and reliable data insights a thing of the present. Details at: https://app.akaocr.io