项目介绍
该项目主要基于TensorRT模型部署套件，在C#平台部署Yolov8模型，包括Yolov8系列的对象检测、图像分割、姿态识别和图像分类模型，实现C#平台推理加速Yolov8模型。
完整范例代码：
GitHub平台：guojin-yan/Csharp_deploy_Yolov8 (github.com)
Gitee平台：Guojin Yan/基于Csharp部署Yolov8系列模型 (gitee.com)
1. OpenVINO™
NVIDIA®TensorRT的核心™ 是一个C++库，有助于在NVIDIA图形处理单元（GPU）上进行高性能推理。TensorRT采用一个经过训练的网络，该网络由一个网络定义和一组经过训练的参数组成，并生成一个高度优化的运行时引擎，为该网络执行推理。TensorRT通过C++和Python提供API，帮助通过网络定义API表达深度学习模型，或通过解析器加载预定义模型，从而使TensorRT能够在NVIDIA GPU上优化和运行它们。TensorRT应用了图优化、层融合等优化，同时还利用高度优化的内核的不同集合找到了该模型的最快实现。TensorRT还提供了一个运行时，您可以使用该运行时在开普勒一代以后的所有NVIDIA GPU上执行该网络。TensorRT还包括Tegra中引入的可选高速混合精度功能™ X1，并用Pascal™, Volta™, Turing™, and NVIDIA® Ampere GPU 架构。
在推理过程中，基于 TensorRT 的应用程序的执行速度可比 CPU 平台的速度快 40 倍。借助 TensorRT，您可以优化在所有主要框架中训练的神经网络模型，精确校正低精度，并最终将模型部署到超大规模数据中心、嵌入式或汽车产品平台中。
2. Yolov8模型
由Ultralytics开发的Ultralytics YOLOv8是一种尖端的，最先进的（SOTA）模型，它建立在以前的YOLO版本成功的基础上，并引入了新功能和改进，以进一步提高性能和灵活性。YOLOv8 设计为快速、准确且易于使用，使其成为各种对象检测、图像分割和图像分类任务的绝佳选择。
YOLOv8是YOLO系列目标检测算法的最新版本，相比于之前的版本，YOLOv8具有更快的推理速度、更高的精度、更加易于训练和调整、更广泛的硬件支持以及原生支持自定义数据集等优势。这些优势使得YOLOv8成为了目前业界最流行和成功的目标检测算法之一。
2.1 安装转换插件
安装ultralytics
pip install ultralytics
安装ONNX
导出ONNX格式模型要求插件
pip install onnx
安装OpenVINO
导出IR模型要求插件
pip install openvino-dev
2.2 获取Yolov8部署模型
Detection
下载模型：
wget https://github.com/ultralytics/assets/releases/download/v0.0.0/yolov8n.pt
模型转换—ONNX
yolo export model=yolov8s.pt imgsz=640 format=onnx opset=12
Segmentation
下载模型：
wget https://github.com/ultralytics/assets/releases/download/v0.0.0/yolov8s-seg.pt
模型转换—ONNX
yolo export model=yolov8s-seg.pt imgsz=640 format=onnx opset=12
Classification
下载模型：
wget https://github.com/ultralytics/assets/releases/download/v0.0.0/yolov8s-cls.pt
模型转换—ONNX
yolo export model=yolov8s-cls.pt imgsz=640 format=onnx opset=12
Pose
下载模型：
wget https://github.com/ultralytics/assets/releases/download/v0.0.0/yolov8s-pose.pt
模型转换—ONNX
yolo export model=yolov8s-pose.pt imgsz=640 format=onnx opset=12
3. TensorRTSharp安装
官方发行的 TensorRT未提供C#编程语言接口，因此在使用时无法实现在C#中利用 TensorRT进行模型部署。在该项目中，利用动态链接库功能，调用官方依赖库，实现在C#中部署深度学习模型。
4.1 TensorRT安装
TensorRT依赖于CUDA加速，因此需要同时安装CUDA与TensorRT才可以使用，且CUDA与TensorRT版本之间需要对应，否者使用会出现较多问题，因此此处并未提供Nuget包，组要根据自己电脑配置悬着合适的版本安装后重新编译本项目源码，下面是TensorRT安装教程：【TensorRT】NVIDIA TensorRT 安装 (Windows C++)_椒颜皮皮虾྅的博客-CSDN博客
4.2 TensorRTSharp 配置
- 获取TensorRTSharp项目源码**
GitHub:
git clone https://github.com/guojin-yan/TensorRTSharp.git
Gitee:
git clone https://gitee.com/guojin-yan/TensorRTSharp.git
- 添加项目引用
TensorRTSharp项目主要包含TensorRTSharpExterm C++ 接口项目和TensorRTSharp C# 类项目，将该项目添加到当前解决中，并增加对TensorRTSharp C# 类项目的引用即可。
由于不同电脑安装的TensorRT和CUDA位置不同，因此在使用中可能会获取不到依赖项，因此建议此处按照TensorRT和CUDA位置重新配置和编译TensorRTSharpExterm C++ 接口项目，C++ 项目配置参考【TensorRT】NVIDIA TensorRT 安装 (Windows C++)_椒颜皮皮虾྅的博客-CSDN博客。
4. Yolov8 detection
4.1 模型推理
基于OpenVINO 和C#同步推理代码的关键片段如下所示：
// 加载推理模型
Nvinfer nvinfer = new Nvinfer(model_path);
// 创建缓存区
nvinfer.creat_gpu_buffer();
// 处理输入数据
Mat image = new Mat(image_path);
int max_image_length = image.Cols > image.Rows ? image.Cols : image.Rows;
Mat max_image = Mat.Zeros(new OpenCvSharp.Size(max_image_length, max_image_length), MatType.CV_8UC3);
Rect roi = new Rect(0, 0, image.Cols, image.Rows);
image.CopyTo(new Mat(max_image, roi));
// 获取缩放比例
float[] factors = new float[2];
factors[0] = factors[1] = (float)(max_image_length / 640.0);
// 获取图片数据
byte[] image_data = max_image.ImEncode(".bmp");
//存储byte的长度
ulong image_size = Convert.ToUInt64(image_data.Length);
// 加载推理图片数据
nvinfer.load_image_data("images", image_data, image_size, BNFlag.Normal);
// 模型推理
nvinfer.infer();
// 读取推理结果
float[] result_array = new float[8400 * 84];
result_array = nvinfer.read_infer_result("output0", 8400 * 84);
// 处理推理结果
DetectionResult result_pro = new DetectionResult(classer_path, factors);
Mat result_image = result_pro.draw_result(result_pro.process_result(result_array), image.Clone());
// 清除推理通道
nvinfer.delete();
4.2 模型推理结果
基于WinForm平台，此处搭建了推理测试平台测试Yolov8 detection模型，结果如图所示：
5. Yolov8 segmentation
5.1 模型推理
基于OpenVINO 和C#同步推理代码的关键片段如下所示：
// 加载推理模型
Nvinfer nvinfer = new Nvinfer(model_path);
// 创建缓存区
nvinfer.creat_gpu_buffer();
// 处理输入数据
Mat image = new Mat(image_path);
int max_image_length = image.Cols > image.Rows ? image.Cols : image.Rows;
Mat max_image = Mat.Zeros(new OpenCvSharp.Size(max_image_length, max_image_length), MatType.CV_8UC3);
Rect roi = new Rect(0, 0, image.Cols, image.Rows);
image.CopyTo(new Mat(max_image, roi));
// 获取缩放比例
float[] factors = new float[4];
factors[0] = factors[1] = (float)(max_image_length / 640.0);
factors[2] = image.Rows;
factors[3] = image.Cols;
// 获取图片数据
byte[] image_data = max_image.ImEncode(".bmp");
//存储byte的长度
ulong image_size = Convert.ToUInt64(image_data.Length);
// 加载推理图片数据
nvinfer.load_image_data("images", image_data, image_size, BNFlag.Normal);
// 模型推理
nvinfer.infer();
// 读取推理结果
float[] det_result_array = new float[8400 * 116];
float[] proto_result_array = new float[32 * 160 * 160];
det_result_array = nvinfer.read_infer_result("output0", 8400 * 116);
proto_result_array = nvinfer.read_infer_result("output1", 32 * 160 * 160);
// 处理推理结果
SegmentationResult result_pro = new SegmentationResult(classer_path, factors);
Mat result_image = result_pro.draw_result(result_pro.process_result(det_result_array, proto_result_array), image.Clone());
// 清除推理通道
nvinfer.delete();
5.2 模型推理结果
基于WinForm平台，此处搭建了推理测试平台测试Yolov8 detection模型，结果如图所示：
6. Yolov8 Classification
6.1 模型推理
基于OpenVINO 和C#同步推理代码的关键片段如下所示：
// 加载推理模型
Nvinfer nvinfer = new Nvinfer(model_path);
// 创建缓存区
nvinfer.creat_gpu_buffer();
// 处理输入数据
Mat image = new Mat(image_path);
int max_image_length = image.Cols > image.Rows ? image.Cols : image.Rows;
Mat max_image = Mat.Zeros(new OpenCvSharp.Size(max_image_length, max_image_length), MatType.CV_8UC3);
Rect roi = new Rect(0, 0, image.Cols, image.Rows);
image.CopyTo(new Mat(max_image, roi));
// 获取图片数据
byte[] image_data = max_image.ImEncode(".bmp");
//存储byte的长度
ulong image_size = Convert.ToUInt64(image_data.Length);
// 加载推理图片数据
nvinfer.load_image_data("images", image_data, image_size, BNFlag.Normal);
// 模型推理
nvinfer.infer();
// 读取推理结果
float[] result_array = new float[1000];
result_array = nvinfer.read_infer_result("output0", 1000);
// 处理推理结果
ClasResult result_pro = new ClasResult(classer_path);
KeyValuePair<string, float> result_cls = result_pro.process_result(result_array);
Mat result_image = result_pro.draw_result(result_cls, image.Clone());
// 清除推理通道
nvinfer.delete();
6.2 模型推理结果
基于WinForm平台，此处搭建了推理测试平台测试Yolov8 Classification模型，结果如图所示：
7. Yolov8 Pose
7.1 模型推理
基于OpenVINO 和C#同步推理代码的关键片段如下所示：
// 加载推理模型
Nvinfer nvinfer = new Nvinfer(model_path);
// 创建缓存区
nvinfer.creat_gpu_buffer();
// 处理输入数据
Mat image = new Mat(image_path);
int max_image_length = image.Cols > image.Rows ? image.Cols : image.Rows;
Mat max_image = Mat.Zeros(new OpenCvSharp.Size(max_image_length, max_image_length), MatType.CV_8UC3);
Rect roi = new Rect(0, 0, image.Cols, image.Rows);
image.CopyTo(new Mat(max_image, roi));
float[] result_array = new float[8400 * 56];
// 获取缩放比例
float[] factors = new float[2];
factors[0] = factors[1] = (float)(max_image_length / 640.0);
// 获取图片数据
byte[] image_data = max_image.ImEncode(".bmp");
//存储byte的长度
ulong image_size = Convert.ToUInt64(image_data.Length);
// 加载推理图片数据
nvinfer.load_image_data("images", image_data, image_size, BNFlag.Normal);
// 模型推理
nvinfer.infer();
// 读取推理结果
float[] result_array = new float[8400 * 56];
result_array = nvinfer.read_infer_result("output0", 8400 * 56);
// 处理推理结果
PoseResult result_pro = new PoseResult( factors);
Mat result_image = result_pro.draw_result(result_pro.process_result(result_array), image.Clone());
// 清除推理通道
nvinfer.delete();
7.2 模型推理结果
基于WinForm平台，此处搭建了推理测试平台测试Yolov8 Pose模型，结果如图所示：
8. 结果处理
8.1 结果类: Result
结果类中主要存放模型预测结果，主要是已经处理后的预测结果，主要包括classes识别类别结果、scores识别分数结果、rects识别框结果、masks语义分割结果、poses关键点结果等五种结果，包含了目标检测、语义分割、姿态识别三种模型的输出结果。
public class Result
{
// 获取结果长度
public int length
{
get
{
return scores.Count;
}
}
// 识别结果类
public List<string> classes = new List<string>();
// 置信值
public List<float> scores = new List<float>();
// 预测框
public List<Rect> rects = new List<Rect>();
// 分割区域
public List<Mat> masks = new List<Mat>();
// 人体关键点
public List<PoseData> poses = new List<PoseData>();
}
为了更具不同的模型增加不同的结果，此处重载了add方法。
public class Result
{
/// <summary>
/// 物体检测
/// </summary>
/// <param name="score">预测分数</param>
/// <param name="rect">识别框</param>
/// <param name="cla">识别类</param>
public void add(float score, Rect rect, string cla)
{
scores.Add(score);
rects.Add(rect);
classes.Add(cla);
}
/// <summary>
/// 物体分割
/// </summary>
/// <param name="score">预测分数</param>
/// <param name="rect">识别框</param>
/// <param name="cla">识别类</param>
/// <param name="mask">语义分割结果</param>
public void add(float score, Rect rect, string cla, Mat mask)
{
scores.Add(score);
rects.Add(rect);
classes.Add(cla);
masks.Add(mask);
}
/// <summary>
/// 关键点预测
/// </summary>
/// <param name="score">预测分数</param>
/// <param name="rect">识别框</param>
/// <param name="pose">关键点数据</param>
public void add(float score, Rect rect, PoseData pose)
{
scores.Add(score);
rects.Add(rect);
poses.Add(pose);
}
}
为了更好的保存关键点数据，在此处定义了一个结构体：PoseData，用于专门存放关键点数据，主要包括关键点以及对应点的分数。
/// <summary>
/// 人体关键点数据
/// </summary>
public struct PoseData
{
public float[] score = new float[17];
public List<Point> point = new List<Point>();
public PoseData(float[] data, float[] scales)
{
for (int i = 0; i < 17; i++)
{
Point p = new Point((int)(data[3 * i] * scales[0]), (int)(data[3 * i + 1] * scales[1]));
this.point.Add(p);
this.score[i] = data[3 * i + 2];
}
}
}
8.2 预测结果处理
结果处理基类：ResultBase
为了处理不同的模型结果，此处定义了一个模型结果处理基类，定义了一些结果处理常用成员以及方法。
public class ResultBase
{
// 识别结果类型
public string[] class_names;
// 图片信息 缩放比例h, 缩放比例h,,height, width
public float[] scales;
// 置信度阈值
public float score_threshold;
// 非极大值抑制阈值
public float nms_threshold;
public ResultBase() { }
/// <summary>
/// 读取本地识别结果类型文件到内存
/// </summary>
/// <param name="path">文件路径</param>
public void read_class_names(string path)
{
List<string> str = new List<string>();
StreamReader sr = new StreamReader(path);
string line;
while ((line = sr.ReadLine()) != null)
{
str.Add(line);
}
class_names = str.ToArray();
}
}
目标识别结果处理：DetectionResult
该方法继承于ResultBase类，主要用于处理目标检测识别结果。
public class DetectionResult : ResultBase
{
/// <summary>
/// 结果处理类构造
/// </summary>
/// <param name="path">识别类别文件地址</param>
/// <param name="scales">缩放比例</param>
/// <param name="score_threshold">分数阈值</param>
/// <param name="nms_threshold">非极大值抑制阈值</param>
public DetectionResult(string path, float[] scales, float score_threshold = 0.25f, float nms_threshold = 0.5f)
{
read_class_names(path);
this.scales = scales;
this.score_threshold = score_threshold;
this.nms_threshold = nms_threshold;
}
/// <summary>
/// 结果处理
/// </summary>
/// <param name="result">模型预测输出</param>
/// <returns>模型识别结果</returns>
public Result process_result(float[] result)
{
Mat result_data = new Mat(84, 8400, MatType.CV_32F, result);
result_data = result_data.T();
// 存放结果list
List<Rect> position_boxes = new List<Rect>();
List<int> class_ids = new List<int>();
List<float> confidences = new List<float>();
// 预处理输出结果
for (int i = 0; i < result_data.Rows; i++)
{
Mat classes_scores = result_data.Row(i).ColRange(4, 84);//GetArray(i, 5, classes_scores);
Point max_classId_point, min_classId_point;
double max_score, min_score;
// 获取一组数据中最大值及其位置
Cv2.MinMaxLoc(classes_scores, out min_score, out max_score,
out min_classId_point, out max_classId_point);
// 置信度 0～1之间
// 获取识别框信息
if (max_score > 0.25)
{
float cx = result_data.At<float>(i, 0);
float cy = result_data.At<float>(i, 1);
float ow = result_data.At<float>(i, 2);
float oh = result_data.At<float>(i, 3);
int x = (int)((cx - 0.5 * ow) * this.scales[0]);
int y = (int)((cy - 0.5 * oh) * this.scales[1]);
int width = (int)(ow * this.scales[0]);
int height = (int)(oh * this.scales[1]);
Rect box = new Rect();
box.X = x;
box.Y = y;
box.Width = width;
box.Height = height;
position_boxes.Add(box);
class_ids.Add(max_classId_point.X);
confidences.Add((float)max_score);
}
}
// NMS非极大值抑制
int[] indexes = new int[position_boxes.Count];
CvDnn.NMSBoxes(position_boxes, confidences, this.score_threshold, this.nms_threshold, out indexes);
Result re_result = new Result();
// 将识别结果绘制到图片上
for (int i = 0; i < indexes.Length; i++)
{
int index = indexes[i];
int idx = class_ids[index];
re_result.add(confidences[index], position_boxes[index], this.class_names[class_ids[index]]);
}
return re_result;
}
/// <summary>
/// 结果绘制
/// </summary>
/// <param name="result">识别结果</param>
/// <param name="image">绘制图片</param>
/// <returns></returns>
public Mat draw_result(Result result, Mat image)
{
// 将识别结果绘制到图片上
for (int i = 0; i < result.length; i++)
{
//Console.WriteLine(result.rects[i]);
Cv2.Rectangle(image, result.rects[i], new Scalar(0, 0, 255), 2, LineTypes.Link8);
Cv2.Rectangle(image, new Point(result.rects[i].TopLeft.X, result.rects[i].TopLeft.Y - 20),
new Point(result.rects[i].BottomRight.X, result.rects[i].TopLeft.Y), new Scalar(0, 255, 255), -1);
Cv2.PutText(image, result.classes[i] + "-" + result.scores[i].ToString("0.00"),
new Point(result.rects[i].X, result.rects[i].Y - 10),
HersheyFonts.HersheySimplex, 0.6, new Scalar(0, 0, 0), 1);
}
return image;
}
}
语义分割结果处理：SegmentationResult
该方法继承于ResultBase类，主要用于处理语义分割识别结果。
public class SegmentationResult : ResultBase
{
/// <summary>
/// 结果处理类构造
/// </summary>
/// <param name="path">识别类别文件地址</param>
/// <param name="scales">缩放比例</param>
/// <param name="score_threshold">分数阈值</param>
/// <param name="nms_threshold">非极大值抑制阈值</param>
public SegmentationResult(string path, float[] scales, float score_threshold = 0.25f, float nms_threshold = 0.5f)
{
read_class_names(path);
this.scales = scales;
this.score_threshold = score_threshold;
this.nms_threshold = nms_threshold;
}
/// <summary>
/// sigmoid函数
/// </summary>
/// <param name="a"></param>
/// <returns></returns>
private float sigmoid(float a)
{
float b = 1.0f / (1.0f + (float)Math.Exp(-a));
return b;
}
/// <summary>
/// 结果处理
/// </summary>
/// <param name="detect">目标检测输出</param>
/// <param name="proto">语义分割输出</param>
/// <returns></returns>
public Result process_result(float[] detect, float[] proto)
{
Mat detect_data = new Mat(116, 8400, MatType.CV_32F, detect);
Mat proto_data = new Mat(32, 25600, MatType.CV_32F, proto);
detect_data = detect_data.T();
// 存放结果list
List<Rect> position_boxes = new List<Rect>();
List<int> class_ids = new List<int>();
List<float> confidences = new List<float>();
List<Mat> masks = new List<Mat>();
// 预处理输出结果
for (int i = 0; i < detect_data.Rows; i++)
{
Mat classes_scores = detect_data.Row(i).ColRange(4, 84);//GetArray(i, 5, classes_scores);
Point max_classId_point, min_classId_point;
double max_score, min_score;
// 获取一组数据中最大值及其位置
Cv2.MinMaxLoc(classes_scores, out min_score, out max_score,
out min_classId_point, out max_classId_point);
// 置信度 0～1之间
// 获取识别框信息
if (max_score > 0.25)
{
Mat mask = detect_data.Row(i).ColRange(84, 116);
float cx = detect_data.At<float>(i, 0);
float cy = detect_data.At<float>(i, 1);
float ow = detect_data.At<float>(i, 2);
float oh = detect_data.At<float>(i, 3);
int x = (int)((cx - 0.5 * ow) * this.scales[0]);
int y = (int)((cy - 0.5 * oh) * this.scales[1]);
int width = (int)(ow * this.scales[0]);
int height = (int)(oh * this.scales[1]);
Rect box = new Rect();
box.X = x;
box.Y = y;
box.Width = width;
box.Height = height;
position_boxes.Add(box);
class_ids.Add(max_classId_point.X);
confidences.Add((float)max_score);
masks.Add(mask);
}
}
// NMS非极大值抑制
int[] indexes = new int[position_boxes.Count];
CvDnn.NMSBoxes(position_boxes, confidences, this.score_threshold, this.nms_threshold, out indexes);
Result re_result = new Result(); // 输出结果类
// 带有颜色的RGB图像
Mat rgb_mask = Mat.Zeros(new Size((int)scales[3], (int)scales[2]), MatType.CV_8UC3);
Random rd = new Random(); // 产生随机数
// 识别结果
for (int i = 0; i < indexes.Length; i++)
{
int index = indexes[i];
// 分割范围
Rect box = position_boxes[index];
int box_x1 = Math.Max(0, box.X);
int box_y1 = Math.Max(0, box.Y);
int box_x2 = Math.Max(0, box.BottomRight.X);
int box_y2 = Math.Max(0, box.BottomRight.Y);
// 分割结果
Mat original_mask = masks[index] * proto_data;
for (int col = 0; col < original_mask.Cols; col++)
{
original_mask.At<float>(0, col) = sigmoid(original_mask.At<float>(0, col));
}
// 1x25600 -> 160x160 转为原始大小
Mat reshape_mask = original_mask.Reshape(1, 160);
//Console.WriteLine("m1.size = {0}", m1.Size());
// 缩放后分割大小
int mx1 = Math.Max(0, (int)((box_x1 / scales[0]) * 0.25));
int mx2 = Math.Max(0, (int)((box_x2 / scales[0]) * 0.25));
int my1 = Math.Max(0, (int)((box_y1 / scales[1]) * 0.25));
int my2 = Math.Max(0, (int)((box_y2 / scales[1]) * 0.25));
// 裁剪分割区域
Mat mask_roi = new Mat(reshape_mask, new OpenCvSharp.Range(my1, my2), new OpenCvSharp.Range(mx1, mx2));
// 将分割区域转换到图片实际大小
Mat actual_maskm = new Mat();
Cv2.Resize(mask_roi, actual_maskm, new Size(box_x2 - box_x1, box_y2 - box_y1));
// 二值化分割区域
for (int r = 0; r < actual_maskm.Rows; r++)
{
for (int c = 0; c < actual_maskm.Cols; c++)
{
float pv = actual_maskm.At<float>(r, c);
if (pv > 0.5)
{
actual_maskm.At<float>(r, c) = 1.0f;
}
else
{
actual_maskm.At<float>(r, c) = 0.0f;
}
}
}
// 预测
Mat bin_mask = new Mat();
actual_maskm = actual_maskm * 200;
actual_maskm.ConvertTo(bin_mask, MatType.CV_8UC1);
if ((box_y1 + bin_mask.Rows) >= scales[2])
{
box_y2 = (int)scales[2] - 1;
}
if ((box_x1 + bin_mask.Cols) >= scales[3])
{
box_x2 = (int)scales[3] - 1;
}
// 获取分割区域
Mat mask = Mat.Zeros(new Size((int)scales[3], (int)scales[2]), MatType.CV_8UC1);
bin_mask = new Mat(bin_mask, new OpenCvSharp.Range(0, box_y2 - box_y1), new OpenCvSharp.Range(0, box_x2 - box_x1));
Rect roi = new Rect(box_x1, box_y1, box_x2 - box_x1, box_y2 - box_y1);
bin_mask.CopyTo(new Mat(mask, roi));
// 分割区域上色
Cv2.Add(rgb_mask, new Scalar(rd.Next(0, 255), rd.Next(0, 255), rd.Next(0, 255)), rgb_mask, mask);
re_result.add(confidences[index], position_boxes[index], this.class_names[class_ids[index]], rgb_mask.Clone());
}
return re_result;
}
/// <summary>
/// 结果绘制
/// </summary>
/// <param name="result">识别结果</param>
/// <param name="image">绘制图片</param>
/// <returns></returns>
public Mat draw_result(Result result, Mat image)
{
Mat masked_img = new Mat();
// 将识别结果绘制到图片上
for (int i = 0; i < result.length; i++)
{
Cv2.Rectangle(image, result.rects[i], new Scalar(0, 0, 255), 2, LineTypes.Link8);
Cv2.Rectangle(image, new Point(result.rects[i].TopLeft.X, result.rects[i].TopLeft.Y - 20),
new Point(result.rects[i].BottomRight.X, result.rects[i].TopLeft.Y), new Scalar(0, 255, 255), -1);
Cv2.PutText(image, result.classes[i] + "-" + result.scores[i].ToString("0.00"),
new Point(result.rects[i].X, result.rects[i].Y - 10),
HersheyFonts.HersheySimplex, 0.6, new Scalar(0, 0, 0), 1);
Cv2.AddWeighted(image, 0.5, result.masks[i], 0.5, 0, masked_img);
}
return masked_img;
}
}
姿态识别结果处理：PoseResult
该方法继承于ResultBase类，主要用于处理姿态识别识别结果。文章来源：https://www.toymoban.com/news/detail-741099.html
public class PoseResult : ResultBase
{
/// <summary>
/// 结果处理类构造
/// </summary>
/// <param name="scales">缩放比例</param>
/// <param name="score_threshold">分数阈值</param>
/// <param name="nms_threshold">非极大值抑制阈值</param>
public PoseResult(float[] scales, float score_threshold = 0.25f, float nms_threshold = 0.5f)
{
this.scales = scales;
this.score_threshold = score_threshold;
this.nms_threshold = nms_threshold;
}
/// <summary>
/// 结果处理
/// </summary>
/// <param name="result">模型预测输出</param>
/// <returns>模型识别结果</returns>
public Result process_result(float[] result)
{
Mat result_data = new Mat(56, 8400, MatType.CV_32F, result);
result_data = result_data.T();
// 存放结果list
List<Rect> position_boxes = new List<Rect>();
List<float> confidences = new List<float>();
List<PoseData> pose_datas = new List<PoseData>();
// 预处理输出结果
for (int i = 0; i < result_data.Rows; i++)
{
// 获取识别框和关键点信息
if (result_data.At<float>(i, 4) > 0.25)
{
//Console.WriteLine(max_score);
float cx = result_data.At<float>(i, 0);
float cy = result_data.At<float>(i, 1);
float ow = result_data.At<float>(i, 2);
float oh = result_data.At<float>(i, 3);
int x = (int)((cx - 0.5 * ow) * this.scales[0]);
int y = (int)((cy - 0.5 * oh) * this.scales[1]);
int width = (int)(ow * this.scales[0]);
int height = (int)(oh * this.scales[1]);
Rect box = new Rect();
box.X = x;
box.Y = y;
box.Width = width;
box.Height = height;
Mat pose_mat = result_data.Row(i).ColRange(5, 56);
float[] pose_data = new float[51];
pose_mat.GetArray<float>(out pose_data);
PoseData pose = new PoseData(pose_data, this.scales);
position_boxes.Add(box);
confidences.Add((float)result_data.At<float>(i, 4));
pose_datas.Add(pose);
}
}
// NMS非极大值抑制
int[] indexes = new int[position_boxes.Count];
CvDnn.NMSBoxes(position_boxes, confidences, this.score_threshold, this.nms_threshold, out indexes);
Result re_result = new Result();
// 将识别结果绘制到图片上
for (int i = 0; i < indexes.Length; i++)
{
int index = indexes[i];
re_result.add(confidences[index], position_boxes[index], pose_datas[i]);
//Console.WriteLine("rect: {0}, score: {1}", position_boxes[index], confidences[index]);
}
return re_result;
}
/// <summary>
/// 结果绘制
/// </summary>
/// <param name="result">识别结果</param>
/// <param name="image">绘制图片</param>
/// <returns></returns>
public Mat draw_result(Result result, Mat image)
{
// 将识别结果绘制到图片上
for (int i = 0; i < result.length; i++)
{
//Console.WriteLine(result.rects[i]);
Cv2.Rectangle(image, result.rects[i], new Scalar(0, 0, 255), 2, LineTypes.Link8);
Cv2.Rectangle(image, new Point(result.rects[i].TopLeft.X, result.rects[i].TopLeft.Y - 20),
new Point(result.rects[i].BottomRight.X, result.rects[i].TopLeft.Y), new Scalar(0, 255, 255), -1);
Cv2.PutText(image, "person -" + result.scores[i].ToString("0.00"),
new Point(result.rects[i].X, result.rects[i].Y - 10),
HersheyFonts.HersheySimplex, 0.6, new Scalar(0, 0, 0), 1);
draw_poses(result.poses[i], ref image);
}
return image;
}
/// <summary>
/// 关键点结果绘制
/// </summary>
/// <param name="pose">关键点数据</param>
/// <param name="image"></param>
public void draw_poses(PoseData pose, ref Mat image)
{
// 连接点关系
int[,] edgs = new int[17, 2] { { 0, 1 }, { 0, 2}, {1, 3}, {2, 4}, {3, 5}, {4, 6}, {5, 7}, {6, 8},
{7, 9}, {8, 10}, {5, 11}, {6, 12}, {11, 13}, {12, 14},{13, 15 }, {14, 16 }, {11, 12 } };
// 颜色库
Scalar[] colors = new Scalar[18] { new Scalar(255, 0, 0), new Scalar(255, 85, 0), new Scalar(255, 170, 0),
new Scalar(255, 255, 0), new Scalar(170, 255, 0), new Scalar(85, 255, 0), new Scalar(0, 255, 0),
new Scalar(0, 255, 85), new Scalar(0, 255, 170), new Scalar(0, 255, 255), new Scalar(0, 170, 255),
new Scalar(0, 85, 255), new Scalar(0, 0, 255), new Scalar(85, 0, 255), new Scalar(170, 0, 255),
new Scalar(255, 0, 255), new Scalar(255, 0, 170), new Scalar(255, 0, 85) };
// 绘制阈值
double visual_thresh = 0.4;
// 绘制关键点
for (int p = 0; p < 17; p++)
{
if (pose.score[p] < visual_thresh)
{
continue;
}
Cv2.Circle(image, pose.point[p], 2, colors[p], -1);
}
// 绘制
for (int p = 0; p < 17; p++)
{
if (pose.score[edgs[p, 0]] < visual_thresh || pose.score[edgs[p, 1]] < visual_thresh)
{
continue;
}
float[] point_x = new float[] { pose.point[edgs[p, 0]].X, pose.point[edgs[p, 1]].X };
float[] point_y = new float[] { pose.point[edgs[p, 0]].Y, pose.point[edgs[p, 1]].Y };
Point center_point = new Point((int)((point_x[0] + point_x[1]) / 2), (int)((point_y[0] + point_y[1]) / 2));
double length = Math.Sqrt(Math.Pow((double)(point_x[0] - point_x[1]), 2.0) + Math.Pow((double)(point_y[0] - point_y[1]), 2.0));
int stick_width = 2;
Size axis = new Size(length / 2, stick_width);
double angle = (Math.Atan2((double)(point_y[0] - point_y[1]), (double)(point_x[0] - point_x[1]))) * 180 / Math.PI;
Point[] polygon = Cv2.Ellipse2Poly(center_point, axis, (int)angle, 0, 360, 1);
Cv2.FillConvexPoly(image, polygon, colors[p]);
}
}
}
分类结果处理：ClasResult
该方法继承于ResultBase类，主要用于处理分类识别结果。文章来源地址https://www.toymoban.com/news/detail-741099.html
public class ClasResult : ResultBase
{
/// <summary>
/// 分类结果构造函数
/// </summary>
/// <param name="path">类别文件</param>
public ClasResult(string path)
{
read_class_names(path);
}
/// <summary>
/// 结果处理
/// </summary>
/// <param name="result">模型输出结果</param>
/// <returns>识别结果与分数</returns>
public KeyValuePair<string, float> process_result(float[] result)
{
int clas = 0;
float score = result[0];
for (int i = 0; i < result.Length; i++)
{
float temp = result[i];
if (score <= temp)
{
score = temp;
clas = i;
}
}
return new KeyValuePair<string, float>(this.class_names[clas], score);
}
/// <summary>
/// 绘制识别结果
/// </summary>
/// <param name="result">识别结果</param>
/// <param name="image"></param>
/// <returns></returns>
public Mat draw_result(KeyValuePair<string, float> result, Mat image)
{
Cv2.PutText(image, result.Key + ": " + result.Value.ToString("0.00"),
new Point(25, 30), HersheyFonts.HersheySimplex, 1, new Scalar(0, 0, 255), 2);
return image;
}
}
到了这里，关于【Yolov8】基于C#和TensorRT部署Yolov8全系列模型的文章就介绍完了。如果您还想了解更多内容，请在右上角搜索TOY模板网以前的文章或继续浏览下面的相关文章，希望大家以后多多支持TOY模板网！
