swift - Core Image 透视校正后 Swift Vision 的矩形检测失败

Question

我正在使用 macOS Big Sur 版本 11.2.3 和 Xcode 版本 12.4。我想获得具有透视失真的数独图像的外部正方形。我通过以下方式进行操作：

1. 执行矩形检测请求。这提供了外部矩形的点。

2. 执行透视校正。这提供了一个完美的二次矩形。

3. 现在我想在数独外框裁剪图像。

4. 对透视校正图像执行第二次矩形检测请求，以获取用于裁剪操作的矩形。

令人惊讶的是，矩形检测结果是一个空数组。

我有一个怀疑，可能是什么原因。

打印出原始 CGImage 的属性提供：

Original image:
    <CGImage 0x7f92e4415560> (IP)
        <<CGColorSpace 0x6000035faf40> (kCGColorSpaceICCBased; kCGColorSpaceModelRGB; sRGB IEC61966-2.1)>
            width = 2448, height = 3264, bpc = 8, bpp = 32, row bytes = 9792
            kCGImageAlphaNoneSkipLast | 0 (default byte order)  | kCGImagePixelFormatPacked
            is mask? No, has masking color? No, has soft mask? No, has matte? No, should interpolate? Yes
    2021-04-06 19:15:04.445374+0200 StackExchangeHilfe[1959:100561] Metal API Validation Enabled

打印出透视校正的 CGImage 的属性提供：

    Corrected image:
    <CGImage 0x7f92f451f180> (DP)
        <<CGColorSpace 0x6000035fae80> (kCGColorSpaceDeviceRGB)>
            width = 2073, height = 2194, bpc = 8, bpp = 32, row bytes = 8320
            kCGImageAlphaPremultipliedLast | 0 (default byte order)  | kCGImagePixelFormatPacked
            is mask? No, has masking color? No, has soft mask? No, has matte? No, should interpolate? Yes

bitmapInfo 有区别：

原始图像：kCGImageAlphaNoneSkipLast

修正图像：kCGImageAlphaPremultipliedLast

任何其他 CIFilter 都不会更改 bitmapInfo。

我试图改变这个值，但它是一个只读变量。但是，也许我的怀疑是完全错误的。

无论如何，有人可以帮忙吗？提前致谢。

    import UIKit
    import Vision

    class ViewController: UIViewController {
        @IBOutlet weak var origImageView: UIImageView!
        @IBOutlet weak var correctedImageView: UIImageView!
        
        let imageName = "sudoku"
        var origImage: UIImage!
        
        override func viewDidLoad() {
            super.viewDidLoad()
            origImage = UIImage(named: imageName)
            origImageView.image = origImage
            let correctedImage = performOperationsWithUIImage(origImage)
            correctedImageView.image = correctedImage
        }
        
        func performOperationsWithUIImage(_ image: UIImage) -> UIImage? {
            let cgImage = image.cgImage!
            print("Original image:")
            print("\(String(describing: cgImage))")
            
            // Create rectangle detect request
            let rectDetectRequest = VNDetectRectanglesRequest()
             // Customize & configure the request to detect only certain rectangles.
            rectDetectRequest.maximumObservations = 8 // Vision currently supports up to 16.
            rectDetectRequest.minimumAspectRatio = 0.8 // height / width
            rectDetectRequest.quadratureTolerance = 30
            rectDetectRequest.minimumSize = 0.5
            rectDetectRequest.minimumConfidence = 0.6
            
            // Create a request handler.
            let imageRequestHandler = VNImageRequestHandler(cgImage: cgImage, orientation: .up, options: [:])
            // Send the requests to the request handler.
            do {
                try imageRequestHandler.perform([rectDetectRequest])
                } catch let error as NSError {
                    print("Failed to perform first image request: \(error)")
                }
            guard let results = rectDetectRequest.results as? [VNRectangleObservation]
            else {return nil}
            print("\nFirst rectangle request result:")
            print("\(results.count) rectangle(s) detected:")
            print("\(String(describing: results))")
            
            // Perform pespective correction
            let width = Int(cgImage.width)
            let height = Int(cgImage.height)
            guard let filter = CIFilter(name:"CIPerspectiveCorrection")  else { return nil }
            
            filter.setValue(CIImage(image: image), forKey: "inputImage")
            filter.setValue(CIVector(cgPoint: VNImagePointForNormalizedPoint(results.first!.topLeft, width, height)), forKey: "inputTopLeft")
            filter.setValue(CIVector(cgPoint: VNImagePointForNormalizedPoint(results.first!.topRight, width, height)), forKey: "inputTopRight")
            filter.setValue(CIVector(cgPoint: VNImagePointForNormalizedPoint(results.first!.bottomLeft, width, height)), forKey: "inputBottomLeft")
            filter.setValue(CIVector(cgPoint: VNImagePointForNormalizedPoint(results.first!.bottomRight, width, height)), forKey: "inputBottomRight")
            
            guard
                let outputCIImage = filter.outputImage,
                let outputCGImage = CIContext(options: nil).createCGImage(outputCIImage, from: outputCIImage.extent)  else {return nil}
            
            print("\nCorrected image:")
            print("\(String(describing: outputCGImage))")
            
            // Perform another rectangle detection
            let newImageRequestHandler = VNImageRequestHandler(cgImage: outputCGImage, orientation: .up, options: [:])
            // Send the requests to the request handler.
            do {
                try newImageRequestHandler.perform([rectDetectRequest])
                } catch let error as NSError {
                    print("Failed to perform second image request: \(error)")
                }
            guard let newResults = rectDetectRequest.results as? [VNRectangleObservation]
            else {return nil}
            print("\nSecond rectangle request result:")
            print("\(newResults.count) rectangle(s) detected:")
            print("\(String(describing: newResults))")

            return UIImage(cgImage: outputCGImage)
        }


    }