如何在VB.NET中模拟ISO噪声以优化神经网络训练数据集？

核心思路 #

模拟高ISO噪声需要贴合真实相机的噪声特性，主要结合两种典型噪声：

• 高斯噪声：模拟传感器的亮度随机噪声，暗部区域噪声会更显著
• 椒盐噪声：模拟像素随机死点或过曝点，比例随ISO等级提升而增加

实现代码 #

以下是可直接集成到数据集生成工具的VB.NET代码：

Imports System.Drawing
Imports System.Drawing.Imaging
Imports System.Math

Public Module IsoNoiseGenerator
    ''' <summary>
    ''' 为图像添加模拟高ISO噪声
    ''' </summary>
    ''' <param name="sourceImage">原始干净图像</param>
    ''' <param name="isoLevel">模拟的ISO等级（建议200-6400，值越大噪声越强）</param>
    ''' <returns>添加噪声后的图像</returns>
    Public Function AddIsoNoise(sourceImage As Image, isoLevel As Integer) As Image
        ' 基于ISO等级计算噪声强度参数
        Dim noiseFactor As Double = Math.Log(isoLevel / 100) * 0.3
        Dim saltPepperRatio As Double = Math.Log(isoLevel / 100) * 0.01

        Dim bmp As New Bitmap(sourceImage)
        Dim rect As New Rectangle(0, 0, bmp.Width, bmp.Height)
        Dim bmpData As BitmapData = bmp.LockBits(rect, ImageLockMode.ReadWrite, bmp.PixelFormat)
        Dim ptr As IntPtr = bmpData.Scan0
        Dim bytes As Integer = Math.Abs(bmpData.Stride) * bmp.Height
        Dim rgbValues As Byte() = New Byte(bytes - 1) {}
        System.Runtime.InteropServices.Marshal.Copy(ptr, rgbValues, 0, bytes)

        Dim rng As New Random()

        ' 遍历所有像素（假设为24位RGB格式）
        For i As Integer = 0 To rgbValues.Length - 1 Step 3
            ' 为每个颜色通道添加高斯噪声，暗部增强噪声强度
            For channel As Integer = 0 To 2
                Dim pixelValue As Integer = rgbValues(i + channel)
                Dim darkBoost As Double = 1 + (255 - pixelValue) / 255 * noiseFactor
                Dim noise As Double = GenerateGaussianNoise(0, noiseFactor * darkBoost)
                Dim newValue As Integer = pixelValue + CInt(noise)

                ' 限制像素值在0-255范围内
                newValue = Math.Max(0, Math.Min(255, newValue))
                rgbValues(i + channel) = CByte(newValue)
            Next

            ' 随机添加椒盐噪声
            If rng.NextDouble() < saltPepperRatio Then
                Dim saltOrPepper As Integer = If(rng.NextDouble() < 0.5, 0, 255)
                rgbValues(i) = CByte(saltOrPepper)
                rgbValues(i + 1) = CByte(saltOrPepper)
                rgbValues(i + 2) = CByte(saltOrPepper)
            End If
        Next

        ' 将处理后的像素数据写回图像
        System.Runtime.InteropServices.Marshal.Copy(rgbValues, 0, ptr, bytes)
        bmp.UnlockBits(bmpData)

        Return bmp
    End Function

    ''' <summary>
    ''' 生成高斯分布随机数（Box-Muller变换）
    ''' </summary>
    Private Function GenerateGaussianNoise(mean As Double, stdDev As Double) As Double
        Dim rng As New Random()
        Dim u1 As Double = rng.NextDouble()
        Dim u2 As Double = rng.NextDouble()
        Dim z0 As Double = Math.Sqrt(-2.0 * Math.Log(u1)) * Math.Cos(2.0 * Math.PI * u2)
        Return z0 * stdDev + mean
    End Function
End Module

使用说明 #

1. 将上述模块添加到你的VB.NET项目中
2. 调用示例：

Dim cleanImage As Image = Image.FromFile("path/to/clean/image.jpg")
Dim noisyImage As Image = IsoNoiseGenerator.AddIsoNoise(cleanImage, 3200) ' 模拟ISO3200噪声
noisyImage.Save("path/to/noisy/image.jpg", ImageFormat.Jpeg)

3. 可根据真实相机的噪声特性，调整noiseFactor和saltPepperRatio的计算逻辑
4. 若需更精准的暗部噪声模拟，可进一步优化darkBoost的计算公式

优化建议 #

• 处理大量图像时，可改用并行遍历提升处理速度
• 可采集真实高ISO噪声样本，通过卷积叠加的方式生成更逼真的噪声
• 针对不同相机型号，单独调整噪声参数以匹配其传感器特性

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