Optimizing Small QR Code Detection on iOS Using OpenCV and AVFoundation

Overview

Long‑pressing an image to recognize a QR code is common on mobile devices, but the QR code must remain visually small. To maintain a high recognition rate for tiny QR codes, the article proposes an optimization that separates detection (using OpenCV) from recognition (using AVFoundation’s CIDetector).

Optimization Scheme

The solution consists of two core pipelines: a detection flow that employs image‑processing algorithms and OpenCV, and a recognition flow that leverages the iOS AVFoundation CIDetector. First, the QR‑code region is detected, then the region undergoes enhancement before recognition.

Detection Flow

The detection process runs several stages to locate a QR code, especially in the lower‑left and lower‑right corners where QR codes are most often placed.

Convert the RGBA image to a single‑channel grayscale image using

OpenCV

cvtColor

.

Apply histogram equalization to clarify contours.

Perform gamma correction to boost contrast.

Clone the bottom‑right 20% of the image, enlarge it 3.5× with resize, and run detect.

Clone the bottom‑left 20% of the image, enlarge it 3.5× with resize, and run detect.

If no core QR‑code features remain intact, conclude that the image contains no QR code.

Only the lower corners are examined because business analysis shows that QR codes are rarely placed in the image centre.

Recognition Flow

After detection, the identified region (often a resized large image) is passed to the recognition step.

The QR‑code area is redrawn into a new bitmap with a 15‑pixel quiet zone to improve detection.

The bitmap is handed to

AVFoundation

CIDetector

for decoding.

The article explains why OpenCV is used only for detection: its recognition rate is lower than that of CIDetector, so the two libraries complement each other.

Code Implementation

Key methods are shown below. The detection code uses OpenCV’s Mat operations, while the conversion between UIImage and Mat is handled via CoreGraphics.

Generate Grayscale Image

CGFloat rows = sourceImage.size.height;
CGFloat cols = sourceImage.size.width;

cv::Mat cvMat(rows, cols, CV_8UC4);
CGColorSpaceRef colorSpace = CGImageGetColorSpace(sourceImage.CGImage);
CGContextRef context = CGBitmapContextCreate(cvMat.data,
                                           cols,
                                           rows,
                                           8,
                                           cvMat.step[0],
                                           colorSpace,
                                           kCGImageAlphaNoneSkipLast | kCGBitmapByteOrderDefault);
if (context == NULL) {
    return cvMat;
}
CGContextDrawImage(context, CGRectMake(0, 0, cols, rows), sourceImage.CGImage);
CGContextRelease(context);
CGColorSpaceRelease(colorSpace);

cv::Mat grayMat;
cv::cvtColor(cvMat, grayMat, cv::COLOR_BGR2GRAY);
return grayMat;

Resize for Corner Detection

CGFloat scale = 3.5f;

cv::Mat copyMat = grayMat.rowRange(grayMat.rows * 0.8, grayMat.rows).colRange(grayMat.cols * 0.5, grayMat.cols).clone();
cv::Mat resizeMat;
cv::resize(copyMat, resizeMat, cv::Size(grayMat.cols * scale * 0.5, grayMat.rows * 0.2 * scale));
NSString *result = [self qrcodeQRCodeForGrayMat:resizeMat];

/// Closed‑eye recognition
if (!result.length) {
    UIImage *image = [self UIImageFromCVMat:resizeMat];
    result = [self readQRCodeWithImage:image detectorAccuracy:CIDetectorAccuracyHigh];
}

if (!result.length) {
    copyMat = grayMat.rowRange(grayMat.rows * 0.8, grayMat.rows).colRange(grayMat.cols * 0.0, grayMat.cols * 0.5).clone();
    cv::resize(copyMat, resizeMat, cv::Size(grayMat.cols * scale * 0.5, grayMat.rows * 0.2 * scale));
    result = [self qrcodeQRCodeForGrayMat:resizeMat];
}
return result;

Recognition

vector<cv::Point> output;
if (output.empty()) {
    return nil;
}
// top‑left
cv::Point point0 = output[0];
// bottom‑right
cv::Point point2 = output[2];

CGRect rect = CGRectMake(point0.x, point0.y, point2.x - point0.x, point2.y - point0.y);
CGRect insetRect = CGRectInset(rect, -15, -15);
if (insetRect.origin.x > 0 && insetRect.origin.y > 0 && insetRect.size.width > 0 && insetRect.size.height > 0) {
    rect = insetRect;
}
UIImage *image = [self UIImageFromCVMat:grayMat];
image = [self drawInRect:rect sourceImage:image];
NSString *result = [self readQRCodeWithImage:image detectorAccuracy:CIDetectorAccuracyLow];
return result;

Mat to UIImage Conversion

CGColorSpaceRef colorSpace;
CGBitmapInfo bitmapInfo;
size_t elemsize = cvMat.elemSize();
if (elemsize == 1) {
    colorSpace = CGColorSpaceCreateDeviceGray;
    bitmapInfo = kCGImageAlphaNone | kCGBitmapByteOrderDefault;
} else {
    colorSpace = CGColorSpaceCreateDeviceRGB;
    bitmapInfo = kCGBitmapByteOrder32Host;
    bitmapInfo |= (elemsize == 4) ? kCGImageAlphaPremultipliedFirst : kCGImageAlphaNone;
}
NSData *data = [NSData dataWithBytes:cvMat.data length:elemsize * cvMat.total()];
CGDataProviderRef provider = CGDataProviderCreateWithCFData((__bridge CFDataRef)data);
CGImageRef imageRef = CGImageCreate(cvMat.cols, cvMat.rows, 8, 8 * cvMat.elemSize(), cvMat.step[0], colorSpace, bitmapInfo, provider, NULL, false, kCGRenderingIntentDefault);
UIImage *finalImage = [UIImage imageWithCGImage:imageRef];
CGImageRelease(imageRef);
CGDataProviderRelease(provider);
CGColorSpaceRelease(colorSpace);
return finalImage;

Benefits

Statistical A/B testing shows a 6.8% increase in recognition speed and a 12% boost in recognition rate compared with the previous solution. In internal tests on complex QR‑code accounts, the optimized pipeline achieved a 100% detection rate.

Future Work

The team plans to add median filtering before applying OTSU binarization, which should improve the quality of binary images and further increase recognition reliability.