iOS+OpenCV扫描文档画质不佳，求优化方案（附处理代码）

I get it—trying to replicate that crisp, professional scan quality from apps like Scannable with OpenCV can feel frustrating. Your current code snippet shows you’re starting with a basic brightness adjustment (cv::multiply), but Scannable relies on a layered pipeline of advanced image processing steps that go way beyond global tweaks. Let’s break down the key techniques you’re missing and how to implement them with OpenCV on iOS:

Key Steps to Match Scannable’s Quality #

1. Adaptive Thresholding (Instead of Fixed Brightness/Contrast) #

Scannable doesn’t just crank up brightness globally—it adjusts to local lighting variations (like shadows or uneven illumination). Replace your fixed multiply operation with adaptive thresholding:

// Convert to grayscale first (critical for thresholding)
cv::Mat grayMat;
cv::cvtColor(originalMat, grayMat, cv::COLOR_RGBA2GRAY);

// Apply adaptive Gaussian thresholding to handle uneven lighting
cv::Mat threshMat;
cv::adaptiveThreshold(
    grayMat, threshMat, 255,
    cv::ADAPTIVE_THRESH_GAUSSIAN_C, cv::THRESH_BINARY,
    11, // Block size (adjust based on document size)
    2   // Constant subtracted from mean
);

2. Automatic Perspective Correction #

Scannable automatically detects document edges and straightens skewed scans. You’ll need to add contour detection and perspective warping:

// Find contours in the thresholded image
std::vector<std::vector<cv::Point>> contours;
std::vector<cv::Vec4i> hierarchy;
cv::findContours(threshMat.clone(), contours, hierarchy, cv::RETR_EXTERNAL, cv::CHAIN_APPROX_SIMPLE);

// Sort contours by area to pick the largest one (your document)
std::sort(contours.begin(), contours.end(), [](const std::vector<cv::Point>& a, const std::vector<cv::Point>& b) {
    return cv::contourArea(a) > cv::contourArea(b);
});

// Approximate the contour to a quadrilateral (document edges)
std::vector<cv::Point> approx;
cv::approxPolyDP(contours[0], approx, cv::arcLength(contours[0], true) * 0.02, true);

// Apply perspective warp to straighten the document
cv::Mat warpMat = cv::getPerspectiveTransform(
    cv::Mat(approx), 
    cv::Mat(std::vector<cv::Point>{
        cv::Point(0,0), 
        cv::Point(originalMat.cols,0), 
        cv::Point(originalMat.cols, originalMat.rows), 
        cv::Point(0, originalMat.rows)
    })
);
cv::Mat warpedMat;
cv::warpPerspective(originalMat, warpedMat, warpMat, originalMat.size());

3. White Balance Correction #

Scannable removes yellowish tints and makes whites look pure. Use the gray world assumption for a simple but effective fix:

cv::Mat floatMat;
originalMat.convertTo(floatMat, CV_32F);

// Calculate average color channels
cv::Scalar avg = cv::mean(floatMat);
float avgGray = (avg[0] + avg[1] + avg[2]) / 3.0;

// Adjust each channel to balance white
floatMat.forEach<cv::Vec4f>([avgGray, avg](cv::Vec4f& pixel, const int* position) -> void {
    pixel[0] *= (avgGray / avg[0]);
    pixel[1] *= (avgGray / avg[1]);
    pixel[2] *= (avgGray / avg[2]);
});

// Convert back to 8-bit and normalize
cv::normalize(floatMat, floatMat, 0, 255, cv::NORM_MINMAX);
floatMat.convertTo(originalMat, CV_8UC4);

4. Noise Reduction (Without Blurring Edges) #

Scannable keeps text edges sharp while reducing grain. Use bilateral filtering instead of Gaussian blur:

cv::Mat denoisedMat;
cv::bilateralFilter(grayMat, denoisedMat, 9, 75, 75);

5. Local Contrast Enhancement with CLAHE #

Instead of global contrast adjustments, use Contrast Limited Adaptive Histogram Equalization (CLAHE) to boost detail in dark areas without overexposing bright ones:

cv::Ptr<cv::CLAHE> clahe = cv::createCLAHE(2.0, cv::Size(8,8));
clahe->apply(grayMat, grayMat);

Putting It All Together #

Run these steps in this order for best results:

1. White balance correction
2. Denoising with bilateral filter
3. Edge detection & perspective correction
4. Adaptive thresholding or CLAHE enhancement

By combining these techniques, you’ll get much closer to the polished, professional scans Scannable produces.

内容的提问来源于stack exchange，提问作者Mital Solanki