使用加速器归一化视差图像

Question

我有一个视差图像，我正在使用下面的示例代码对其进行标准化，但它非常慢。 我需要使用一些加速器，如自定义 CIFilter 或任何其他技术，但我不知道怎么做？我目前正在使用 CIContext() 运行代码，它正在 CPU 上运行（不确定）。有没有办法在 GPU 上运行它并在没有自定义 CIfilter 的情况下加速？ 这是当前代码：

extension CVPixelBuffer {

  func normalize() {

    let width = CVPixelBufferGetWidth(self)
    let height = CVPixelBufferGetHeight(self)

    CVPixelBufferLockBaseAddress(self, CVPixelBufferLockFlags(rawValue: 0))

    let baseAddr = CVPixelBufferGetBaseAddress(self)!
    let floatBuffer = unsafeBitCast(CVPixelBufferGetBaseAddress(self), to: UnsafeMutablePointer<Float>.self)

    var minPixel: Float = 1.0
    var maxPixel: Float = 0.0

    for y in 0 ..< height {
      for x in 0 ..< width {
        let pixel = floatBuffer[y * width + x]
        minPixel = min(pixel, minPixel)
        maxPixel = max(pixel, maxPixel)
      }
    }


    let range = maxPixel - minPixel

    for y in 0 ..< height {
      for x in 0 ..< width {
        let pixel = floatBuffer[y * width + x]
        floatBuffer[y * width + x] = (pixel - minPixel) / range
      }
    }

    CVPixelBufferUnlockBaseAddress(self, CVPixelBufferLockFlags(rawValue: 0))
  }
}

Answer 1

您将像素值设为Float 值，因此您也可以使用 vDSP。

vDSP_minv 和 vDSP_maxv 计算极值，并且：

floatBuffer[y * width + x] = (pixel - minPixel) / range

可以替换为vDSP_vasm（您需要乘以range 的倒数）。

查看执行此计算的 vDSP_normalize 也可能有用：

            m = sum(A[n], 0 <= n < N) / N;
            d = sqrt(sum(A[n]**2, 0 <= n < N) / N - m**2);

            if (C)
            {
                // Normalize.
                for (n = 0; n < N; ++n)
                    C[n] = (A[n] - m) / d;
            }

Answer 2

对于您的用例，vImage 可能是最佳选择。请参阅this answer 中的选项 3。

Core Image 中也有一些方法可以做到这一点。我会想象使用CIAreaMinMax 过滤器获取极值，然后使用一些巧妙的混合进行标准化。如果你愿意，我可以详细说明。

Answer 3

我使用 Accelerate Framework vDSP 矢量函数来标准化视差。在 gitHub 中查看修改后的 PhotoBrowse 以获得工作演示。

这是两个函数中的相关代码

extension CVPixelBuffer {
    func vectorNormalize( targetVector: UnsafeMutableBufferPointer<Float>) -> [Float] {
        // range = max - min
        // normalized to 0..1 is (pixel - minPixel) / range

        // see Documentation "Using vDSP for Vector-based Arithmetic" in vDSP under system "Accelerate" documentation

        // see also the Accelerate documentation section 'Vector extrema calculation'
        // Maximium static func maximum<U>(U) -> Float
        //      Returns the maximum element of a single-precision vector.

        //static func minimum<U>(U) -> Float
        //      Returns the minimum element of a single-precision vector.


        let maxValue = vDSP.maximum(targetVector)
        let minValue = vDSP.minimum(targetVector)

        let range = maxValue - minValue
        let negMinValue = -minValue

        let subtractVector = vDSP.add(negMinValue, targetVector)
            // adding negative value is subtracting
        let result = vDSP.divide(subtractVector, range)

        return result
    }

    func setUpNormalize() -> CVPixelBuffer {
        // grayscale buffer float32 ie Float
        // return normalized CVPixelBuffer

        CVPixelBufferLockBaseAddress(self,
                                     CVPixelBufferLockFlags(rawValue: 0))
        let width = CVPixelBufferGetWidthOfPlane(self, 0)
        let height = CVPixelBufferGetHeightOfPlane(self, 0)
        let count = width * height

        let bufferBaseAddress = CVPixelBufferGetBaseAddressOfPlane(self, 0)
            // UnsafeMutableRawPointer

        let pixelBufferBase  = unsafeBitCast(bufferBaseAddress, to: UnsafeMutablePointer<Float>.self)

        let depthCopy  =   UnsafeMutablePointer<Float>.allocate(capacity: count)
        depthCopy.initialize(from: pixelBufferBase, count: count)
        let depthCopyBuffer = UnsafeMutableBufferPointer<Float>(start: depthCopy, count: count)

        let normalizedDisparity = vectorNormalize(targetVector: depthCopyBuffer)

        pixelBufferBase.initialize(from: normalizedDisparity, count: count)
            // copy back the normalized map into the CVPixelBuffer

        depthCopy.deallocate()
//        depthCopyBuffer.deallocate()

        CVPixelBufferUnlockBaseAddress(self, CVPixelBufferLockFlags(rawValue: 0))

        return self

    }

}