说明
官网地址:github.com/microsoft/D…
效果

模型信息
Model Properties
-------------------------
metadata:{}
---------------------------------------------------------------
Inputs
-------------------------
name:input
tensor:Float[-1, 3, 512, 512]
---------------------------------------------------------------
Outputs
-------------------------
name:output
tensor:Float[-1, 3, 512, 512]
---------------------------------------------------------------
项目

代码
using Microsoft.ML.OnnxRuntime
using Microsoft.ML.OnnxRuntime.Tensors
using OpenCvSharp
using System
using System.Collections.Generic
using System.Drawing
using System.Drawing.Imaging
using System.Linq
using System.Windows.Forms
namespace Onnx_Demo
{
public partial class Form1 : Form
{
// ----- 法线估计专用字段 -----
string fileFilter = "*.*|*.bmp;*.jpg;*.jpeg;*.tiff;*.png"
string image_path = ""
string startupPath
DateTime dt1 = DateTime.Now
DateTime dt2 = DateTime.Now
string model_path
Mat image
Mat normal_color_map
SessionOptions options
InferenceSession onnx_session
Tensor<float> input_tensor
List<NamedOnnxValue> input_container
IDisposableReadOnlyCollection<DisposableNamedOnnxValue> result_infer
int inpHeight = 512, inpWidth = 512
public Form1()
{
InitializeComponent()
}
// ----- 按钮1:选择图片 -----
private void button1_Click(object sender, EventArgs e)
{
OpenFileDialog ofd = new OpenFileDialog()
ofd.Filter = fileFilter
if (ofd.ShowDialog() != DialogResult.OK) return
pictureBox1.Image = null
image_path = ofd.FileName
pictureBox1.Image = new Bitmap(image_path)
textBox1.Text = ""
image = new Mat(image_path)
pictureBox2.Image = null
normal_color_map = null
}
// ----- 按钮2:执行法线估计推理 -----
private void button2_Click(object sender, EventArgs e)
{
if (string.IsNullOrEmpty(image_path))
{
MessageBox.Show("请先选择图片!")
return
}
button2.Enabled = false
pictureBox2.Image = null
textBox1.Text = ""
Application.DoEvents()
// 读取原始图像(BGR)
image = new Mat(image_path)
int originalWidth = image.Cols
int originalHeight = image.Rows
// ------------------ 预处理 ------------------
// 1. 缩放至模型输入尺寸 512x512
Mat resized = new Mat()
Cv2.Resize(image, resized, new OpenCvSharp.Size(inpWidth, inpHeight))
// 2. 转换为浮点型并归一化到 [0,1]
resized.ConvertTo(resized, MatType.CV_32FC3, 1.0 / 255.0)
// 3. 分离 BGR 通道,并按 RGB 顺序填充(模型预期 RGB)
Mat[] channels = Cv2.Split(resized)
int channelSize = inpHeight * inpWidth
float[] inputData = new float[3 * channelSize]
// 将 B,G,R 重新排列为 R,G,B
for (int c = 0
{
float[] channelData = new float[channelSize]
System.Runtime.InteropServices.Marshal.Copy(channels[c].Data, channelData, 0, channelSize)
int targetIndex = (c == 0) ? 2 : (c == 2) ? 0 : 1
Array.Copy(channelData, 0, inputData, targetIndex * channelSize, channelSize)
}
// 4. 创建输入张量
input_tensor = new DenseTensor<float>(inputData, new[] { 1, 3, inpHeight, inpWidth })
input_container.Clear()
input_container.Add(NamedOnnxValue.CreateFromTensor("input", input_tensor))
// ------------------ 推理 ------------------
dt1 = DateTime.Now
result_infer = onnx_session.Run(input_container)
dt2 = DateTime.Now
// 获取输出
var output = result_infer.First(x => x.Name == "output").AsTensor<float>()
var dimensions = output.Dimensions.ToArray()
int outChannels = dimensions[1]
int outH = dimensions[2]
int outW = dimensions[3]
float[] normalFloat = output.ToArray()
// 创建三通道法线 Mat (CV_32FC3),形状 H×W×3
Mat normalRaw = new Mat(outH, outW, MatType.CV_32FC3)
// 注意:输出张量布局为 [N,C,H,W],需要转换为 H,W,C 存储
int planeSize = outH * outW
for (int c = 0
{
float[] channelData = new float[planeSize]
Array.Copy(normalFloat, c * planeSize, channelData, 0, planeSize)
// 将每个通道的数据填充到 Mat 的对应通道
Mat channelMat = new Mat(outH, outW, MatType.CV_32FC1)
System.Runtime.InteropServices.Marshal.Copy(channelData, 0, channelMat.Data, planeSize)
// 将单通道合并到 normalRaw
Mat[] destChannels = Cv2.Split(normalRaw)
channelMat.CopyTo(destChannels[c])
Cv2.Merge(destChannels, normalRaw)
}
// ------------------ 后处理 ------------------
// 1. 双线性插值恢复原始尺寸
Mat normalResized = new Mat()
Cv2.Resize(normalRaw, normalResized, new OpenCvSharp.Size(originalWidth, originalHeight), interpolation: InterpolationFlags.Linear)
// 2. 归一化法线向量(确保每个像素的向量长度为1,防止模型输出未严格归一化)
Mat[] normChannels = Cv2.Split(normalResized)
Mat normSq = new Mat()
Cv2.Pow(normChannels[0], 2, normSq)
Mat tmp = new Mat()
Cv2.Pow(normChannels[1], 2, tmp)
Cv2.Add(normSq, tmp, normSq)
Cv2.Pow(normChannels[2], 2, tmp)
Cv2.Add(normSq, tmp, normSq)
Mat norm = new Mat()
Cv2.Sqrt(normSq, norm)
norm += 1e-8
for (int i = 0
{
Cv2.Divide(normChannels[i], norm, normChannels[i])
}
Mat normalizedNormal = new Mat()
Cv2.Merge(normChannels, normalizedNormal)
// 3. 将法线从 [-1,1] 映射到 [0,255] 并转为 8UC3 用于显示
Mat normalDisplay = new Mat()
normalizedNormal.ConvertTo(normalDisplay, MatType.CV_32FC3, 127.5, 127.5)
normalDisplay.ConvertTo(normalDisplay, MatType.CV_8UC3)
// 注意:OpenCV 默认 BGR 顺序,而法线 RGB 直接显示可能会颜色偏差,若需要保持 RGB 可交换 R 和 B
// 这里为了视觉效果,交换 R 和 B 通道使显示更自然(法线常见可视化中 R 对应 X,G 对应 Y,B 对应 Z)
Mat[] displayChannels = Cv2.Split(normalDisplay)
// 交换 R 和 B
Mat temp = displayChannels[0].Clone()
displayChannels[0] = displayChannels[2]
displayChannels[2] = temp
Cv2.Merge(displayChannels, normalDisplay)
normal_color_map = normalDisplay.Clone()
// 显示结果
pictureBox2.Image = new Bitmap(normal_color_map.ToMemoryStream())
textBox1.Text = $"推理耗时: {(dt2 - dt1).TotalMilliseconds:F2} ms";
button2.Enabled = true;
}
// ----- 按钮3:保存法线彩色图 -----
private void button3_Click(object sender, EventArgs e)
{
if (normal_color_map == null || normal_color_map.Empty())
{
MessageBox.Show("请先执行法线估计!");
return;
}
SaveFileDialog sdf = new SaveFileDialog();
sdf.Title = "保存法线彩色图";
sdf.Filter = "PNG图片 (*.png)|*.png|JPEG图片 (*.jpg)|*.jpg|BMP图片 (*.bmp)|*.bmp";
sdf.FilterIndex = 1;
if (sdf.ShowDialog() == DialogResult.OK)
{
Cv2.ImWrite(sdf.FileName, normal_color_map);
MessageBox.Show($"保存成功: {sdf.FileName}");
}
}
// ----- 窗体加载:初始化 ONNX 模型 -----
private void Form1_Load(object sender, EventArgs e)
{
startupPath = Application.StartupPath;
// 法线估计模型路径(请根据实际位置修改)
model_path = System.IO.Path.Combine(startupPath, "model", "normal-model-vitb16_384.onnx");
if (!System.IO.File.Exists(model_path))
{
MessageBox.Show($"模型文件不存在: {model_path}\n请将模型放置于 {startupPath}\model\ 目录下");
return;
}
options = new SessionOptions();
options.LogSeverityLevel = OrtLoggingLevel.ORT_LOGGING_LEVEL_INFO;
options.AppendExecutionProvider_CPU(0);
// 若需 CUDA,可取消注释
// options.AppendExecutionProvider_CUDA(0);
onnx_session = new InferenceSession(model_path, options);
input_container = new List<NamedOnnxValue>();
// 可选默认测试图片
string testImg = System.IO.Path.Combine(startupPath, "test_img", "0.jpg");
if (System.IO.File.Exists(testImg))
{
image_path = testImg;
pictureBox1.Image = new Bitmap(image_path);
image = new Mat(image_path);
}
}
// ----- 双击图片放大(保留原功能,假设存在 Common 类)-----
private void pictureBox1_DoubleClick(object sender, EventArgs e)
{
Common.ShowNormalImg(pictureBox1.Image);
}
private void pictureBox2_DoubleClick(object sender, EventArgs e)
{
Common.ShowNormalImg(pictureBox2.Image);
}
}
}