基于置信区间的宽松式Python二维码检测实现方案问询

问题背景 #

我需在拍摄远离镜头人物的视频中，检测其身上附着的二维码。尝试过OpenCV与pyzbar的多种实现，但准确率仅约50%。裁剪图像可提升检测效果，但二维码随人物移动，需复杂鲁棒的裁剪逻辑，超出预期工作量。

核心需求为检测二维码是否存在，检测位置优先级高于解码，且仅需在少量视频帧中完成解码。目前无法理解OpenCV中QRCodeDetector().detect()的实现逻辑，希望找到基于置信区间的宽松检测方法（不限OpenCV或pyzbar）。当前实现阈值未知，存在帧间二维码位置/姿态未变却无法检测的情况，需放宽检测要求。

视频帧示例：二维码在待搜索帧中占比相对较小，附着于远离镜头的人物身上。

当前实现代码 #

video = cv2.VideoCapture(input_video)
ret, frame = video.read()

while ret:
    ret, frame = video.read()

    qrCodeDetector = cv2.QRCodeDetector()
    points = qrCodeDetector.detect(image)[1]

    if points is not None:
        points = points[0]

        # 计算二维码中心
        center = tuple(np.mean(np.array(points),axis=0).astype(int))

        # 绘制二维码中心标记
        cv2.circle(frame,center,50,color=(255,0,0),thickness=2)

        cv2.imshow('frame', frame)
        cv2.waitKey(1)
video.release()
cv2.destroyAllWindows()

可行的宽松检测方案 #

1. 调整OpenCV QR检测参数与预处理 #

OpenCV的QRCodeDetector可通过setEpsX和setEpsY调整检测阈值，数值越大检测越宽松；同时配合图像预处理提升小二维码辨识度：

import cv2
import numpy as np

video = cv2.VideoCapture(input_video)
ret, frame = video.read()

# 初始化检测器并调整参数
qr_detector = cv2.QRCodeDetector()
qr_detector.setEpsX(0.1)
qr_detector.setEpsY(0.1)

while ret:
    ret, frame = video.read()
    if not ret:
        break
        
    # 图像预处理：灰度化+降噪+对比度增强
    gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
    blurred = cv2.GaussianBlur(gray, (3,3), 0)
    enhanced = cv2.equalizeHist(blurred)
    
    # 检测二维码
    points = qr_detector.detect(enhanced)[1]
    if points is not None:
        points = points[0]
        center = tuple(np.mean(np.array(points),axis=0).astype(int))
        cv2.circle(frame, center, 50, (255,0,0), 2)
        
    cv2.imshow('frame', frame)
    cv2.waitKey(1)

video.release()
cv2.destroyAllWindows()

2. pyzbar配合图像缩放与宽松匹配 #

通过放大图像提升小二维码的识别率，同时不严格校验解码结果（优先检测存在性）：

import cv2
from pyzbar.pyzbar import decode, ZBarSymbol

video = cv2.VideoCapture(input_video)
ret, frame = video.read()
scale_factor = 2  # 缩放倍数，根据实际情况调整

while ret:
    ret, frame = video.read()
    if not ret:
        break
        
    # 放大图像并灰度化
    resized = cv2.resize(frame, None, fx=scale_factor, fy=scale_factor, interpolation=cv2.INTER_CUBIC)
    gray_resized = cv2.cvtColor(resized, cv2.COLOR_BGR2GRAY)
    
    # 仅检测QR码，不强制校验解码内容
    results = decode(gray_resized, symbols=[ZBarSymbol.QRCODE])
    for result in results:
        rect = result.rect
        # 转回原图像坐标
        center = (rect.left + rect.width//2, rect.top + rect.height//2)
        center = (center[0]//scale_factor, center[1]//scale_factor)
        cv2.circle(frame, center, 50, (255,0,0), 2)
        
    cv2.imshow('frame', frame)
    cv2.waitKey(1)

video.release()
cv2.destroyAllWindows()

3. 模板匹配+跟踪检测 #

利用已成功检测到的二维码作为模板，后续帧通过模板匹配跟踪，适合姿态变化不大的场景：

import cv2
import numpy as np

def get_qr_template(frame, points):
    # 提取二维码ROI并做透视变换生成模板
    points = np.array(points, dtype=np.float32)
    target_size = (200, 200)
    target_points = np.array([[0,0], [target_size[0],0], [target_size[0],target_size[1]], [0,target_size[1]]], dtype=np.float32)
    M = cv2.getPerspectiveTransform(points, target_points)
    template = cv2.warpPerspective(frame, M, target_size)
    return cv2.cvtColor(template, cv2.COLOR_BGR2GRAY)

video = cv2.VideoCapture(input_video)
ret, frame = video.read()
qr_detector = cv2.QRCodeDetector()
qr_detector.setEpsX(0.1)
qr_detector.setEpsY(0.1)
template = None
match_threshold = 0.5  # 匹配阈值，越低检测越宽松

while ret:
    ret, frame = video.read()
    if not ret:
        break
        
    gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
    # 模板匹配跟踪
    if template is not None:
        res = cv2.matchTemplate(gray, template, cv2.TM_CCOEFF_NORMED)
        loc = np.where(res >= match_threshold)
        for pt in zip(*loc[::-1]):
            center = (pt[0] + template.shape[1]//2, pt[1] + template.shape[0]//2)
            cv2.circle(frame, center, 50, (255,0,0), 2)
    
    # 原检测逻辑，用于更新模板
    points = qr_detector.detect(gray)[1]
    if points is not None:
        template = get_qr_template(frame, points[0])
        
    cv2.imshow('frame', frame)
    cv2.waitKey(1)

video.release()
cv2.destroyAllWindows()

4. YOLO目标检测模型 #

使用预训练或自定义训练的YOLO二维码检测模型，通过调整置信度阈值控制检测宽松程度，对小目标、姿态变化鲁棒性更强：

import cv2
from ultralytics import YOLO

video = cv2.VideoCapture(input_video)
ret, frame = video.read()
# 加载预训练二维码检测模型（可自行训练或使用公开模型）
model = YOLO('yolov8n-qr.pt')
conf_threshold = 0.2  # 置信度阈值，越低检测越宽松

while ret:
    ret, frame = video.read()
    if not ret:
        break
        
    # 推理检测二维码
    results = model(frame, conf=conf_threshold)
    for result in results:
        for box in result.boxes:
            x1, y1, x2, y2 = map(int, box.xyxy[0])
            center = ((x1+x2)//2, (y1+y2)//2)
            cv2.circle(frame, center, 50, (255,0,0), 2)
        
    cv2.imshow('frame', frame)
    cv2.waitKey(1)

video.release()
cv2.destroyAllWindows()

内容的提问来源于stack exchange，提问作者Kunal Shah