C# 代码示例

本章会详细介绍DaoAI World SDK中包含的C#代码示例。

引入库 

在C#示例中，我们引入了以下几个库，其中 DaoAI.DeepLearningCLI 是用于引入DaoAI World SDK的库。

using System;
using System.Collections.Generic;
using System.Linq;
using System.Text;
using System.Drawing;
using System.Threading.Tasks;
using DaoAI.DeepLearningCLI;

读取图片 

DaoAI World SDK 的模型预测函数需要将图片表示为一维数组（1D array）。以下是从文件中读取图片的代码部分：

// Test image
String root_directory = System.IO.Directory.GetCurrentDirectory();

System.Drawing.Bitmap image = new
    System.Drawing.Bitmap("C:\\Users\\daoai\\Downloads\\DW_SDK\\DW_SDK Example\\Data\\maskrcnn_data\\daoai_1.png"); //图片文件路径
System.Drawing.Bitmap image_copy = new System.Drawing.Bitmap(image);

byte[] pixels = new byte[image.Width * image.Height * 3];
for (int i = 0; i < image.Height; i++)
{
    for (int j = 0; j < image.Width; j++)
    {
        System.Drawing.Color color = image.GetPixel(j, i);
        pixels[(i * image.Width + j) * 3] = (byte)(color.R);
        pixels[(i * image.Width + j) * 3 + 1] = (byte)(color.G);
        pixels[(i * image.Width + j) * 3 + 2] = (byte)(color.B);
    }
}

这里做了一个图片的深度拷贝,然后用 DaoAI.DeepLearningCLI.Image 函数初始化图像对象以便后续使用：

DaoAI.DeepLearningCLI.Image img = new DaoAI.DeepLearningCLI.Image(image.Height, image.Width, DaoAI.DeepLearningCLI.Image.Type.RGB, pixels);
DaoAI.DeepLearningCLI.Image img_copy = img.clone();
byte[] image_data = img_copy.data;

for (int i = 0; i < image.Width; i++)
{
    for (int j = 0; j < image.Height; j++)
    {
        int index_r = i * image.Height + j;
        int index_g = i * image.Height + j + image.Width * image.Height;
        int index_b = i * image.Height + j + 2 * image.Width * image.Height;
        byte r = image_data[index_r];
        byte g = image_data[index_g];
        byte b = image_data[index_b];
        System.Drawing.Color color = Color.FromArgb(r, g, b);
        image_copy.SetPixel(i, j, color);
    }
}

加载深度学习模型 

String data_path = "..\\..\\..\\..\\Data\\";
String model_path = data_path + "model.dwm";
// init model
DaoAI.DeepLearningCLI.Vision.KeypointDetection model = new DaoAI.DeepLearningCLI.Vision.KeypointDetection(model_path);

注意，这里每一个检测任务都有对应的对象：

//实例分割
DaoAI.DeepLearningCLI.Vision.InstanceSegmentation model(model_path) = new DaoAI.DeepLearningCLI.Vision.InstanceSegmentation(model_path);

//关键点检测
DaoAI.DeepLearningCLI.Vision.KeypointDetection model(model_path) = new DaoAI.DeepLearningCLI.Vision.KeypointDetection(model_path);

//图像分类
DaoAI.DeepLearningCLI.Vision.Classification model(model_path) = new DaoAI.DeepLearningCLI.Vision.Classification(model_path);

//目标检测
DaoAI.DeepLearningCLI.Vision.ObjectDetection model(model_path) = new DaoAI.DeepLearningCLI.Vision.ObjectDetection(model_path);

//非监督缺陷检测
DaoAI.DeepLearningCLI.Vision.UnsupervisedDefectSegmentation model(model_path) = new DaoAI.DeepLearningCLI.Vision.AnomalyDetection(model_path);

//监督缺陷检测
DaoAI.DeepLearningCLI.Vision.SupervisedDefectSegmentation model(model_path) = new DaoAI.DeepLearningCLI.Vision.SemanticSegmentation(model_path);

//OCR
DaoAI.DeepLearningCLI.Vision.OCR model(model_path) = new DaoAI.DeepLearningCLI.Vision.OCR(model_path);

//定位模型 (只在工业版支持)
DaoAI.DeepLearningCLI.Vision.Positioning model(model_path) = new DaoAI.DeepLearningCLI.Vision.Positioning(model_path);

//漏错装检测 (只在工业版支持)
DaoAI.DeepLearningCLI.Vision.PresenceChecking model(model_path) = new DaoAI.DeepLearningCLI.Vision.PresenceChecking(model_path);

使用深度学习模型进行预测 

这里定义了置信度阈值(CONFIDENT_THRESHOLD)为 0.5, 并调用 model.inferece() 函数来使用模型进行推理，再使用 .toJSONString()方法打印为 json

Dictionary<DaoAI.DeepLearningCLI.PostProcessType, object> post_params = new Dictionary<DaoAI.DeepLearningCLI.PostProcessType, object>();
post_params[DaoAI.DeepLearningCLI.PostProcessType.CONFIDENT_THRESHOLD] = 0.5;

Console.WriteLine(model.inference(img, post_params).toJSONString());

后处理参数 

模型的预测可以接受后处理参数：

DaoAI.DeepLearningCLI.PostProcessType.CONFIDENCE_THRESHOLD:

置信度阈值，会过滤掉结果中置信度低于设定值的结果

post_params[DaoAI.DeepLearningCLI.PostProcessType.IOU_THRESHOLD:

IOU阈值，会过滤掉结果中IOU低于设定值的结果

post_params[DaoAI.DeepLearningCLI.PostProcessType.SENSITIVITY_THRESHOLD:

非监督缺陷分割（异常检测）模型中使用敏感度，控制模型对于缺陷的敏感度，越高则模型会检测出越多的缺陷，但是容易误检

Dictionary<DaoAI.DeepLearningCLI.PostProcessType, object> post_params = new Dictionary<DaoAI.DeepLearningCLI.PostProcessType, object>();
post_params[DaoAI.DeepLearningCLI.PostProcessType.CONFIDENCE_THRESHOLD] = 0.5; //置信度阈值，会过滤掉结果中置信度低于设定值的结果
post_params[DaoAI.DeepLearningCLI.PostProcessType.IOU_THRESHOLD] = 0.5; //IOU阈值，会过滤掉结果中IOU低于设定值的结果

Console.WriteLine(model.inference(img, post_params));

获取预测结果 

Box（边界框）

Box 是模型预测结果中的边界框，可通过以下方式获取：

// 获取预测结果
DaoAI.DeepLearningCLI.Vision.KeypointDetectionResult prediction = model.Inference(daoaiImage);

// 访问边界框坐标
double x1 = prediction.boxes[0].x1();  // 左上角 x 坐标
double y1 = prediction.boxes[0].y1();  // 左上角 y 坐标
double x2 = prediction.boxes[0].x2();  // 右下角 x 坐标
double y2 = prediction.boxes[0].y2();  // 右下角 y 坐标

Mask（掩码）

Mask 是预测结果中目标物体的外轮廓信息，可以通过以下方式提取多边形区域的顶点：

// 获取预测结果
DaoAI.DeepLearningCLI.Vision.KeypointDetectionResult prediction = model.Inference(daoaiImage);

// 获取多边形区域的第一个顶点
double x = prediction.masks[0].toPolygons()[0].points[0].x; // 顶点的 X 坐标
double y = prediction.masks[0].toPolygons()[0].points[0].y; // 顶点的 Y 坐标

`masks[0]`：提取第一个目标物体的掩码。
`toPolygons()`：将掩码转换为多边形外轮廓。
`points[0]`：获取多边形的第一个顶点坐标。

通过顺序连接多边形的所有顶点，可以形成完整的轮廓区域，用于目标检测、区域分析等应用场景。所有顶点按顺序连接，形成目标物体的完整多边形区域。此方法适合需要详细外轮廓信息的场景，如区域分析或精细标注。

可视化输出 

生成并保存预测结果与原图叠加的可视化图像：

DaoAI.DeepLearningCLI.Image result = DaoAI.DeepLearningCLI.Utils.visualize(img, prediction);
result.save(root_directory + "daoai_output.png");

返回结果示例 

以下是实例分割模型预测后返回的结果示例。

这个结果展示了预测的数量，标签名称，置信度，以及预测框，和多边形掩膜。

{
    "Number of detecions": 1,
    "Detections": [
        {
            "Label": "zheng",
            "Confidence": 0.9523001313209534,
            "Box": [
                955.1925659179688, 316.0162048339844, 1064.072021484375,
                426.4408264160156
            ],
            "Mask": [
                [990.0, 316.0],
                [988.0, 318.0],
                [987.0, 318.0],
                [985.0, 320.0],
                [982.0, 320.0],
                [980.0, 322.0],
                [979.0, 322.0],
                [974.0, 327.0],
                [972.0, 327.0],
                [972.0, 328.0],
                [1040.0, 316.0]
            ]
        }
    ],
    "ImageHeight": 1200,
    "ImageWidth": 1920
}