ios - CoreGraphics - 无法将 UnsafeMutablePointer 替换为 UnsafeMutablePointer

Question

我刚开始弄乱图像处理，我遇到了几个非常奇怪的问题，或者至少我认为是这样。我假设我犯了一些非常愚蠢的错误。

我本来打算就此发布另一个问题，但是使用下面的代码有时我也会得到随机噪声，而不是用户抽取数字的像素表示。如果有人能告诉我为什么会发生这种情况，我将不胜感激。我很难找出原因，因为我阅读的所有内容都表明这段代码应该有效。

如果有人需要更多信息，请告诉我!提前感谢您的帮助!

目标:
首先，获取用户在屏幕上绘制的数字。然后，将图像大小调整为 28 x 28。接下来，将图像转换为灰度并获得归一化像素值数组。最后，将归一化的灰度像素值输入机器学习算法。

[注意:在下图中，点代表 0 值，1 代表值 > 0。]

下面代码的输出效果很好。如果用户画了一个“3”，我通常会得到如下结果:

问题:

如果我将 UnsafeMutablePointer 和 Buffer 的大小更改为 UInt8，我会得到看起来像随机噪声的东西。或者，如果我用 [UInt32](repeating: 0, count: totalBytes) 甚至 [UInt8](repeating: 0, count: totalBytes) 替换 UnsafeMutablePointer 和 Buffer pixel最后变成了0，我真的不明白。

如果我将 UnsafeMutablePointer 和 Buffer 的大小更改为 UInt8，则像素的输出如下:

获取灰度像素的代码:

public extension UIImage
{
    private func grayScalePixels() -> UnsafeMutableBufferPointer<UInt32>?
    {
        guard let cgImage = self.cgImage else { return nil }

        let bitsPerComponent = 8
        let width = cgImage.width
        let height = cgImage.height
        let totalBytes = (width * height)
        let colorSpace = CGColorSpaceCreateDeviceGray()
        let data = UnsafeMutablePointer<UInt32>.allocate(capacity: totalBytes)
        defer { data.deallocate(capacity: totalBytes) }

        guard let imageContext = CGContext(data: data, width: width, height: height, bitsPerComponent: bitsPerComponent, bytesPerRow: width, space: colorSpace, bitmapInfo: 0) else { return nil }
        imageContext.draw(cgImage, in: CGRect(origin: CGPoint.zero, size: CGSize(width: width, height: height)))

        return UnsafeMutableBufferPointer<UInt32>(start: data, count: totalBytes)
    }

    public func normalizedGrayScalePixels() -> [CGFloat]?
    {
        guard let cgImage = self.cgImage else { return nil }
        guard let pixels = self.grayScalePixels() else { return nil }

        let width = cgImage.width
        let height = cgImage.height
        var result = [CGFloat]()

        for y in 0..<height
        {
            for x in 0..<width
            {
                let index = ((width * y) + x)
                let pixel = (CGFloat(pixels[index]) / 255.0)
                result.append(pixel)
            }
        }

        return result
    }
}

取号代码:

    func drawLineFrom(fromPoint: CGPoint, toPoint: CGPoint)
    {
        UIGraphicsBeginImageContextWithOptions(self.view.bounds.size, false, 1)

        self.tempImageView.image?.draw(at: CGPoint.zero)

        let context = UIGraphicsGetCurrentContext()
        context?.move(to: fromPoint)
        context?.addLine(to: toPoint)
        context?.setLineCap(.round)
        context?.setLineWidth(self.brushWidth)
        context?.setStrokeColor(gray: 0, alpha: 1)
        context?.strokePath()

        self.tempImageView.image = UIGraphicsGetImageFromCurrentImageContext()
        self.tempImageView.alpha = self.opacity

        UIGraphicsEndImageContext()
    }

    override func touchesBegan(_ touches: Set<UITouch>, with event: UIEvent?)
    {
        self.swiped = false

        if let touch = touches.first {
            self.lastPoint = touch.location(in: self.view)
        }
    }

    override func touchesMoved(_ touches: Set<UITouch>, with event: UIEvent?)
    {
        self.swiped = true

        if let touch = touches.first
        {
            let currentPoint = touch.location(in: self.view)
            self.drawLineFrom(fromPoint: self.lastPoint, toPoint: currentPoint)

            self.lastPoint = currentPoint
        }
    }

    override func touchesEnded(_ touches: Set<UITouch>, with event: UIEvent?)
    {
        if !swiped {
            self.drawLineFrom(fromPoint: self.lastPoint, toPoint: self.lastPoint)
        }

        self.predictionLabel.text = "Predication: \(self.predict())"

        self.tempImageView.image = nil
    }

预测数字的代码:

    private func printNumber(rowSize: Int, inputs: Vector)
    {
        for (index, pixel) in inputs.enumerated()
        {
            if index % rowSize == 0 { print() }

            if (pixel > 0) {
                print("1", terminator: " ")
            }
            else { print(".", terminator: " ") }
        }

        print()
    }

    private func predict() -> Scalar
    {
        let resizedImaege = self.tempImageView.image!.resizedImage(CGSize(width: 28, height: 28), interpolationQuality: .high)
        let inputs = resizedImaege!.normalizedGrayScalePixels()!.flatMap({ Scalar($0) })
        self.feedforwardResult = self.neuralNetwork!.feedForward(inputs: inputs)

        self.printNumber(rowSize: 28, inputs: inputs)

        let max = self.feedforwardResult!.activations.last!.max()!
        let prediction = self.feedforwardResult!.activations.last!.index(of: max)!
        return Scalar(prediction)
    }

Answer 1

您的代码中有一件非常糟糕的事情是这一行:

    defer { data.deallocate(capacity: totalBytes) }

data.deallocate(capacity: totalBytes) 在退出方法 grayScalePixels() 之前执行。因此，返回的 UnsafeMutableBufferPointer 的 baseAddress 指向一个已经释放的区域，这意味着您在访问该区域时不能期望任何可预测的结果。

如果你想使用UnsafeMutableBufferPointer，你需要在完成对它的所有访问后释放该区域(下面代码中的#1):

private func grayScalePixels() -> UnsafeMutableBufferPointer<UInt8>? {
    guard let cgImage = self.cgImage else { return nil }
    
    let bitsPerComponent = 8
    let width = cgImage.width
    let height = cgImage.height
    let totalBytes = width * height
    let colorSpace = CGColorSpaceCreateDeviceGray()
    let data = UnsafeMutablePointer<UInt8>.allocate(capacity: totalBytes)
    data.initialize(to: UInt8.max, count: totalBytes)   //<- #4
    
    guard let imageContext = CGContext(data: data, width: width, height: height, bitsPerComponent: bitsPerComponent, bytesPerRow: width, space: colorSpace, bitmapInfo: 0) else { return nil }
    imageContext.draw(cgImage, in: CGRect(x: 0, y: 0, width: width, height: height))
    
    return UnsafeMutableBufferPointer(start: data, count: totalBytes)
}

public func normalizedGrayScalePixels() -> [CGFloat]? {
    guard let cgImage = self.cgImage else { return nil }
    guard let pixels = self.grayScalePixels() else { return nil }
    
    let width = cgImage.width
    let height = cgImage.height
    var result: [CGFloat] = []
    
    for y in 0..<height {
        for x in 0..<width {
            let index = width * y + x
            let pixel = CGFloat(pixels[index]) / CGFloat(UInt8.max)
            result.append(pixel)
        }
    }
    pixels.baseAddress!.deinitialize(count: pixels.count)   //<- #2
    pixels.baseAddress!.deallocate(capacity: pixels.count)  //<- #1
    
    return result
}

(#2) 在当前的 Swift 实现中，UInt8 可能不需要 deinitialize，但顺序:allocate - initialize - deinitilize - deallocate 是推荐的方式.

(其他一些行只是我的偏好，并不重要。)

---

否则，如果你想使用 Swift Array 而不是 UnsafeMutableBufferPointer，你可以这样写:

private func grayScalePixels() -> [UInt8]? {
    guard let cgImage = self.cgImage else { return nil }
    
    let bitsPerComponent = 8
    let width = cgImage.width
    let height = cgImage.height
    let totalBytes = width * height
    let colorSpace = CGColorSpaceCreateDeviceGray()
    var byteArray: [UInt8] = Array(repeating: UInt8.max, count: totalBytes) //<- #4
    let success = byteArray.withUnsafeMutableBufferPointer {(buffer)->Bool in
        guard let imageContext = CGContext(data: buffer.baseAddress!, width: width, height: height, bitsPerComponent: bitsPerComponent, bytesPerRow: width, space: colorSpace, bitmapInfo: 0) else { return false }
        imageContext.draw(cgImage, in: CGRect(x: 0, y: 0, width: width, height: height))
        return true;
    }
    return success ? byteArray : nil
}

public func normalizedGrayScalePixels() -> [CGFloat]? {
    guard let cgImage = self.cgImage else { return nil }
    guard let pixels = self.grayScalePixels() else { return nil }
    
    let width = cgImage.width
    let height = cgImage.height
    var result: [CGFloat] = []
    
    for y in 0..<height {
        for x in 0..<width {
            let index = width * y + x
            let pixel = CGFloat(pixels[index]) / CGFloat(UInt8.max)
            result.append(pixel)
        }
    }
    
    return result
}

---

您可能需要修改我上面的代码，使它们适用于您的代码，因为我无法使用您的 UInt32 版本的 grayScalePixels() 重现相同的结果。

---

编辑

我在我的代码中发现了一个问题。您的绘图代码绘制线:

    context?.setStrokeColor(gray: 0, alpha: 1)

灰度 0，黑色。在我的旧代码中，我将位图初始化为:

    data.initialize(to: 0, count: totalBytes)

或:

    var byteArray: [UInt8] = Array(repeating: 0, count: totalBytes)

所以，在黑色上画黑色，结果:全黑，8 位灰度，全 0。(我第一次写 initialize 可能不需要，但那是一个错误。带有 alpha 的图像将与初始位图内容混合绘制。)

我更新的代码(用 #4 标记)用白色初始化位图(8 位灰度，255 == 0xFF == UInt8.max).

而且您最好通过更新 printNumber(rowSize:inputs:) 来检测非白色像素:

private func printNumber(rowSize: Int, inputs: Vector) {
    for (index, pixel) in inputs.enumerated() {
        if index % rowSize == 0 { print() }
        
        if pixel < 1.0 { //<- #4
            print("1", terminator: "")
        }
        else { print(".", terminator: "") }
    }
    
    print()
}

在归一化灰度为 float 时，1.0 是白色的值，您最好将非白色显示为 1。 (或者，找到另一个更好的阈值。)