RAIDI 2025.3.11

推理:

创建检测器

using RAIDI;
using SixLabors.ImageSharp;

// 使用CPU
using var predictor = new Detector(model);
// 使用GPU
using var predictor = new Detector(model, SessionOptions.MakeSessionOptionWithCudaProvider(0));

设置检测器的参数

// 设置交并比(分类、检测、分割）
predictor.Parameters.IoU = .45f
// 设置置信度(分类、检测、分割）
predictor.Parameters.Confidence = .5f
// 使用原始宽高比处理(分类、检测、分割）
predictor.Parameters.ProcessWithOriginalAspectRatio = true
// 分割结果的掩膜置信度
predictor.Parameters.SegmentPixelsWithDefaultConfidence = .5f;

支持的图像格式

(Image image) using SixLabors.ImageSharp.Image (string path) path to image (byte[] data) byte (Stream stream) stream

使用以下代码执行快速推理获取结果(不区分深度学习方法，但编译器可能无法在编译时检查某些属性是否存在）

var result = predictor.Inference("path/to/image");

Console.WriteLine(result);

使用以下代码推理.rairun 分类 模型

IClassificationResult result = predictor.Classify("path/to/image");
// 异步
var result = await predictor.ClassifyAsync("path/to/image");

Console.WriteLine(result);

使用以下代码推理.rairun 目标检测 模型

IDetectionResult result = predictor.Detect("path/to/image");
// 异步
var result = await predictor.DetectAsync("path/to/image");

Console.WriteLine(result);

使用以下代码推理.rairun 实例分割 模型

ISegmentationResult result = predictor.Segment("path/to/image");
// 异步
var result = await predictor.SegmentAsync("path/to/image");

Console.WriteLine(result);

使用以下代码推理.rairun 异常检测 模型

IAnomalyResult result = predictor.Anomaly("path/to/image");
// 异步
var result = await predictor.AnomalyAsync("path/to/image");

Console.WriteLine(result);

绘图

分类、检测、分割

using RAIDI;
using RAIDI.Plotting;
using SixLabors.ImageSharp;

var imagePath = "path/to/image";

using var predictor = new Detector("path/to/model");

var result = await predictor.SegmentAsync(imagePath);

using var image = Image.Load(imagePath);
using var ploted = await result.PlotImageAsync(image);

ploted.Save("./Image.jpg")

异常检测

using RAIDI;

var imagePath = "path/to/image";

using var predictor = new Detector("path/to/model");

// result内包含热力图、合并图、原图图像、平均异常分数、最大异常分数、异常分数的95百分位值、超过预定阈值（128）的像素比例、每个像素的原始异常分数（分数已经归一化到0-255范围）
var result = await predictor.AnomalyAsync(imagePath);

// 使用 ToString 将保存图像及结果
Console.WriteLine(result)

其中RAIDI.Plotting可以设置绘图的字体位置、线宽、颜色等信息。

以下是一个完整的示例

using RAIDI;
using RAIDI.Plotting;
using System.Diagnostics;
using SixLabors.ImageSharp;
using Microsoft.ML.OnnxRuntime;

var output = "./path/to";

if (Directory.Exists(output) == false)
	Directory.CreateDirectory(output);


if (OperatingSystem.IsWindows())
{
	Process.Start(new ProcessStartInfo
	{
		FileName = Path.GetFullPath(output),
		UseShellExecute = true,
	});
}


await Classify(new string[] { "./path/to/Part_A.jpg", "./path/to/Part_B.jpg" }, "./path/to/Sample_Part_best.rairun");

await Detect("./path/to/Abnormal.jpg", "./path/to/Sample_Abnormal_best.rairun");

await Segment("./path/to/Workpiece.jpg", "./assets/to/Sample_Workpiece_best.rairun");

await Anomaly("./path/to/Car.jpg", "./assets/to/Sample_Car_best.rairun");


async Task Classify(string[] images, string model)
{
	Console.WriteLine("\n================ 图像分类 ================\n");

	Console.WriteLine("加载模型...");

	// 使用GPU
	using var predictor = new Detector(model, SessionOptions.MakeSessionOptionWithCudaProvider(0));

	foreach (var image in images)
	{
		Console.WriteLine("执行计算... ({0})", image);
		var result = await predictor.ClassifyAsync(image);

		Console.WriteLine($"结果: {result}");
		Console.WriteLine($"速度: {result.Speed}");

		Console.WriteLine("绘图及保存...");
		using var origin = Image.Load(image);

		using var ploted = await result.PlotImageAsync(origin);

		var pathToSave = Path.Combine(output, Path.GetFileName(image));

		ploted.Save(pathToSave);
	}
}


async Task Detect(string image, string model)
{
	Console.WriteLine("\n================ 目标检测 ================\n");

	Console.WriteLine("加载模型...");
	using var predictor = new Detector(model);

	Console.WriteLine("执行计算...");
	var result = await predictor.DetectAsync(image);

	Console.WriteLine($"结果: {result}");
	Console.WriteLine($"速度: {result.Speed}");

	Console.WriteLine("绘图及保存...");
	using var origin = Image.Load(image);

	using var ploted = await result.PlotImageAsync(origin);

	var pathToSave = Path.Combine(output, Path.GetFileName(image));

	ploted.Save(pathToSave);
}


async Task Segment(string image, string model)
{
	Console.WriteLine("\n================ 实例分割 ================\n");

	Console.WriteLine("加载模型...");
	using var predictor = new Detector(model);

	Console.WriteLine("执行计算...");
	var result = await predictor.SegmentAsync(image);

	Console.WriteLine($"结果: {result}");
	Console.WriteLine($"速度: {result.Speed}");

	Console.WriteLine("绘图及保存...");
	using var origin = Image.Load(image);

	//掩膜置信度设置为0.5，掩膜轮廓线宽设置为0.1
	using var ploted = await result.PlotImageAsync(origin, new SegmentationPlottingOptions { MaskConfidence = 5f, ContoursThickness = 1f });

	var filename = $"{Path.GetFileNameWithoutExtension(image)}_seg";
	var extension = Path.GetExtension(image);

	var pathToSave = Path.Combine(output, filename + extension);

	ploted.Save(pathToSave);
}


async Task Anomaly(string image, string model)
{
	Console.WriteLine("\n================ 异常检测 ================\n");

	Console.WriteLine("加载模型...");
	using var predictor = new Detector(model);

	Console.WriteLine("执行计算...");
	var result = await predictor.AnomalyAsync(image);

	// 保存图像及结果
	Console.WriteLine($"结果: {result}");
	Console.WriteLine($"速度: {result.Speed}");
}