inference.v0.1.960x544.py

import torch
import cv2
import numpy as np
import onnxruntime as ort
import time

# --- CONFIGURATION ---
INPUT_WIDTH = 960
INPUT_HEIGHT = 544
MODEL_PATH = f"meiki.text.detect.v0.1.{INPUT_WIDTH}x{INPUT_HEIGHT}.onnx"
INPUT_IMAGE_PATH = f"input.jpg"
OUTPUT_IMAGE_PATH = f"output.{INPUT_WIDTH}x{INPUT_HEIGHT}.jpg"

# A threshold to filter out weak detections.
# You can adjust this value (e.g., lower to 0.3 for more boxes,
# or raise to 0.5 for fewer, more confident boxes).
CONFIDENCE_THRESHOLD = 0.4

def resize(image: np.ndarray, w, h):
    original_height, original_width, _ = image.shape

    # Calculate the ratio to resize the image.
    ratio_w = w / original_width
    ratio_h = h / original_height

    # Resize the image
    resized_image = cv2.resize(image, (w, h), interpolation=cv2.INTER_LINEAR)

    return resized_image, ratio_w, ratio_h

def main():
    """
    Main function to run the inference process.
    """
    # --- 1. Load the Model ---
    try:
        # Create an inference session with the ONNX model.
        session = ort.InferenceSession(MODEL_PATH, providers=['CUDAExecutionProvider'])
        print("Session providers:", session.get_providers())
        print(f"Successfully loaded model: {MODEL_PATH}")
    except Exception as e:
        print(f"Error: Failed to load the ONNX model. Make sure '{MODEL_PATH}' exists.")
        print(f"Details: {e}")
        return

    # --- 2. Load and Pre-process the Input Image ---
    try:
        # Read the input image from the file. It will be in BGR format by default.
        original_image = cv2.imread(INPUT_IMAGE_PATH)
        if original_image is None:
            raise FileNotFoundError(f"Image not found at '{INPUT_IMAGE_PATH}'")
        print(f"Successfully loaded image: {INPUT_IMAGE_PATH}")
    except Exception as e:
        print(f"Error: {e}")
        return

    resized_image, ratio_w, ratio_h = resize(original_image, INPUT_WIDTH,INPUT_HEIGHT)

    # Normalize the image data to be between 0 and 1.
    img_normalized = resized_image.astype(np.float32) / 255.0

    # The model expects the channel dimension to be first (Channels, Height, Width).
    # OpenCV loads images as (Height, Width, Channels), so we transpose the axes.
    img_transposed = np.transpose(img_normalized, (2, 0, 1))

    image_input_tensor = np.expand_dims(img_transposed, axis=0)

    # --- 3. Run Inference ---
    # The model requires a second input specifying the image size
    sizes_input_tensor = np.array([[INPUT_WIDTH, INPUT_HEIGHT]], dtype=np.int64)

    # Get the names of the model's input nodes.
    input_names = [inp.name for inp in session.get_inputs()]

    # Prepare the dictionary of inputs for the model.
    inputs = {
        input_names[0]: image_input_tensor,
        input_names[1]: sizes_input_tensor
    }

    # Run the model.
    # This model returns three separate outputs: labels, boxes, and confidence scores.
    for i in range(10):
        start = time.perf_counter()
        outputs = session.run(None, inputs)
        print(f"runtime {time.perf_counter() - start}")
    labels, boxes, scores = outputs

    # --- 4. Post-process and Draw Bounding Boxes ---
    # The outputs have an extra batch dimension, so we remove it.
    boxes = boxes[0]
    scores = scores[0]

    print(f"Model returned {len(boxes)} boxes. Filtering with confidence > {CONFIDENCE_THRESHOLD}...")

    # Create a copy of the original image to draw on.
    output_image = original_image.copy()

    # Iterate through the boxes and their corresponding scores.
    confident_boxes_count = 0
    for box, score in zip(boxes, scores):
        # Only process boxes with a confidence score above our threshold.
        if score > CONFIDENCE_THRESHOLD:
            confident_boxes_count += 1
            # The coordinates from the model are relative to the 640x640 padded image.
            # We need to scale them back to the original image's coordinate space.
            x_min, y_min, x_max, y_max = box

            final_x_min = int(x_min / ratio_w)
            final_y_min = int(y_min / ratio_h)
            final_x_max = int(x_max / ratio_w)
            final_y_max = int(y_max / ratio_h)

            # Draw a green rectangle on the output image.
            cv2.rectangle(output_image, (final_x_min, final_y_min), (final_x_max, final_y_max), (0, 255, 0), 2)

    print(f"Found {confident_boxes_count} confident boxes.")

    # --- 5. Save the Final Image ---
    cv2.imwrite(OUTPUT_IMAGE_PATH, output_image)
    print(f"Successfully saved result to: {OUTPUT_IMAGE_PATH}")


if __name__ == "__main__":
    main()