【发布时间】:2020-04-20 04:23:57
【问题描述】:
我有一个使用 tensorflow 2.0 / Keras 构建的模型。输入是具有 28x28 和 1 个通道的图像。该模型被保存并转换为 .tflite 并在我的 swift ios 应用程序中使用。不幸的是,当调用解释器时,我得到的预测与预期截然不同。当我进一步调查时,似乎我的图像准备可能是错误的。以下是我在将像素阵列输入模型之前采取的步骤。
- 加载图像。
- 将图像转换为灰度
- 通过除以 255 来归一化像素值。
1
252
255
253
255
255
255
253
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
253
255
253
254
253
255
254
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
255
248
255
253
255
255
248
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
255
255
255
255
253
255
255
252
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
255
250
254
255
248
253
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
255
255
255
251
255
255
254
253
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
250
255
255
255
255
248
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
255
255
255
255
253
255
251
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
这是我的代码
let im = UIImage(named: "dotsgray")!
let i = (im.pixelBufferGray(width: 28, height: 28))!
i.normalize()
extension UIImage {
public func pixelBufferGray(width: Int, height: Int) -> CVPixelBuffer? {
return pixelBuffer(width: width, height: height,
pixelFormatType: kCVPixelFormatType_OneComponent8,
colorSpace: CGColorSpaceCreateDeviceGray(),
alphaInfo: .none)
}
func pixelBuffer(width: Int, height: Int, pixelFormatType: OSType,
colorSpace: CGColorSpace, alphaInfo: CGImageAlphaInfo) -> CVPixelBuffer? {
var maybePixelBuffer: CVPixelBuffer?
let attrs = [kCVPixelBufferCGImageCompatibilityKey: kCFBooleanTrue,
kCVPixelBufferCGBitmapContextCompatibilityKey: kCFBooleanTrue]
let status = CVPixelBufferCreate(kCFAllocatorDefault,
width,
height,
pixelFormatType,
attrs as CFDictionary,
&maybePixelBuffer)
guard status == kCVReturnSuccess, let pixelBuffer = maybePixelBuffer else {
return nil
}
CVPixelBufferLockBaseAddress(pixelBuffer, CVPixelBufferLockFlags(rawValue: 0))
let pixelData = CVPixelBufferGetBaseAddress(pixelBuffer)
guard let context = CGContext(data: pixelData,
width: width,
height: height,
bitsPerComponent: 8,
bytesPerRow: CVPixelBufferGetBytesPerRow(pixelBuffer),
space: colorSpace,
bitmapInfo: alphaInfo.rawValue)
else {
return nil
}
UIGraphicsPushContext(context)
context.translateBy(x: 0, y: CGFloat(height))
context.scaleBy(x: 1, y: -1)
self.draw(in: CGRect(x: 0, y: 0, width: width, height: height))
UIGraphicsPopContext()
CVPixelBufferUnlockBaseAddress(pixelBuffer, CVPixelBufferLockFlags(rawValue: 0))
return pixelBuffer
}
}
extension CVPixelBuffer {
func normalize() {
// 1
let bytesPerRow = CVPixelBufferGetBytesPerRow(self)
let totalBytes = CVPixelBufferGetDataSize(self)
let width = bytesPerRow / MemoryLayout<UInt8>.size
let height = totalBytes / bytesPerRow
// 2
CVPixelBufferLockBaseAddress(self, CVPixelBufferLockFlags(rawValue: 0))
// 3
let floatBuffer = unsafeBitCast(
CVPixelBufferGetBaseAddress(self),
to: UnsafeMutablePointer<Double>.self)
// 4
var minPixel: Double = 1.0
var maxPixel: Double = 0.0
// 5
for i in 0 ..< width * height {
let pixel = floatBuffer[i]
minPixel = min(pixel, minPixel)
maxPixel = max(pixel, maxPixel)
}
// 6
let range = maxPixel - minPixel
// 7
for i in 0 ..< width * height {
let pixel = floatBuffer[i]
floatBuffer[i] = (pixel - minPixel) / range
}
// 8
CVPixelBufferUnlockBaseAddress(self, CVPixelBufferLockFlags(rawValue: 0))
}
【问题讨论】:
-
检查
CVPixelBufferGetBytesPerRow(pixelBuffer)的值 -
CVPixelBufferGetBytesPerRow(self) 为 64,CVPixelBufferGetDataSize(self) 为 1792
-
您应该编写代码来执行 RGB -> 灰度转换并在不涉及像素缓冲区和 CoreVideo API 的情况下测试该逻辑。您需要同时处理色彩空间转换和伽玛。然后,一旦您的代码经过全面测试并且已知可以正常工作,请将该逻辑逐行应用于 CoreVideo 提供的像素。有关 RGB -> Grayscale 的详细信息,请参阅此问题(还提供源代码)stackoverflow.com/questions/53911662/…
标签: swift rgb pixel tensorflow2.0 grayscale