使用 YOLO 和 EasyOCR 从视频文件中检测车牌

本文将介绍如何通过Python中的YOLO（ou Only Look Once）和EasyOCR（光学字符识别）技术来实现从视频文件中检测车牌。本技术依托于深度学习，以实现车牌的即时检测与识别。

从视频文件中检测车牌

先决条件

在我们开始之前，请确保已安装以下Python包：

复制

pip install opencv-python ultralytics easyocr Pillow numpy

实施步骤

步骤1：初始化库

我们将首先导入必要的库。我们将使用OpenCV进行视频处理，使用YOLO进行目标检测，并使用EasyOCR读取检测到的车牌上的文字。

复制

import cv2
from ultralytics import YOLO
import easyocr
from PIL import Image
import numpy as np

# Initialize EasyOCR reader
reader = easyocr.Reader(['en'], gpu=False)

# Load your YOLO model (replace with your model's path)
model = YOLO('best_float32.tflite', task='detect')

# Open the video file (replace with your video file path)
video_path = 'sample4.mp4'
cap = cv2.VideoCapture(video_path)

# Create a VideoWriter object (optional, if you want to save the output)
output_path = 'output_video.mp4'
fourcc = cv2.VideoWriter_fourcc(*'mp4v')
out = cv2.VideoWriter(output_path, fourcc, 30.0, (640, 480))  # Adjust frame size if necessary

步骤2：处理视频帧

我们将从视频文件中读取每一帧，处理它以检测车牌，然后应用OCR来识别车牌上的文字。为了提高性能，我们可以跳过每第三帧的处理。

复制

# Frame skipping factor (adjust as needed for performance)
frame_skip = 3  # Skip every 3rd frame
frame_count = 0

while cap.isOpened():
    ret, frame = cap.read()  # Read a frame from the video
    if not ret:
        break  # Exit loop if there are no frames left

    # Skip frames
    if frame_count % frame_skip != 0:
        frame_count += 1
        continue  # Skip processing this frame

    # Resize the frame (optional, adjust size as needed)
    frame = cv2.resize(frame, (640, 480))  # Resize to 640x480

    # Make predictions on the current frame
    results = model.predict(source=frame)

    # Iterate over results and draw predictions
    for result in results:
        boxes = result.boxes  # Get the boxes predicted by the model
        for box in boxes:
            class_id = int(box.cls)  # Get the class ID
            confidence = box.conf.item()  # Get confidence score
            coordinates = box.xyxy[0]  # Get box coordinates as a tensor

            # Extract and convert box coordinates to integers
            x1, y1, x2, y2 = map(int, coordinates.tolist())  # Convert tensor to list and then to int

            # Draw the box on the frame
            cv2.rectangle(frame, (x1, y1), (x2, y2), (0, 255, 0), 2)  # Draw rectangle

            # Try to apply OCR on detected region
            try:
                # Ensure coordinates are within frame bounds
                r0 = max(0, x1)
                r1 = max(0, y1)
                r2 = min(frame.shape[1], x2)
                r3 = min(frame.shape[0], y2)

                # Crop license plate region
                plate_region = frame[r1:r3, r0:r2]

                # Convert to format compatible with EasyOCR
                plate_image = Image.fromarray(cv2.cvtColor(plate_region, cv2.COLOR_BGR2RGB))
                plate_array = np.array(plate_image)

                # Use EasyOCR to read text from plate
                plate_number = reader.readtext(plate_array)
                concat_number = ' '.join([number[1] for number in plate_number])
                number_conf = np.mean([number[2] for number in plate_number])

                # Draw the detected text on the frame
                cv2.putText(
                    img=frame,
                    text=f"Plate: {concat_number} ({number_conf:.2f})",
                    org=(r0, r1 - 10),
                    fontFace=cv2.FONT_HERSHEY_SIMPLEX,
                    fontScale=0.7,
                    color=(0, 0, 255),
                    thickness=2
                )

            except Exception as e:
                print(f"OCR Error: {e}")
                pass

    # Show the frame with detections
    cv2.imshow('Detections', frame)

    # Write the frame to the output video (optional)
    out.write(frame)

    if cv2.waitKey(1) & 0xFF == ord('q'):
        break  # Exit loop if 'q' is pressed

    frame_count += 1  # Increment frame count

# Release resources
cap.release()
out.release()  # Release the VideoWriter object if used
cv2.destroyAllWindows()

代码解释：

• 启动EasyOCR：设置EasyOCR以识别英文字符。
• 导入YOLO模型：从特定路径载入YOLO模型，需替换为模型的实际路径。
• 视频帧读取：利用OpenCV打开视频文件，若需保存输出，则启动VideoWriter。
• 帧尺寸调整与处理：逐帧读取并调整尺寸，随后使用模型预测车牌位置。
• 绘制识别结果：在视频帧上标出识别到的车牌边界框，并裁剪出车牌区域以进行OCR识别。
• 执行OCR：EasyOCR识别裁剪后的车牌图像中的文本，并在帧上展示识别结果及置信度。
• 视频输出：处理后的视频帧可显示在窗口中，也可以选择保存为视频文件。

结论

本段代码展示了如何综合运用YOLO和EasyOCR技术，从视频文件中检测并识别车牌。遵循这些步骤，你可以为自己的需求构建相似的系统。根据实际情况，灵活调整参数和优化模型性能。