C# Onnx实现轻量实时的M-LSD直线检测
作者:天天代码码天天
介绍
github地址:https://github.com/navervision/mlsd
M-LSD: Towards Light-weight and Real-time Line Segment Detection
Official Tensorflow implementation of "M-LSD: Towards Light-weight and Real-time Line Segment Detection" (AAAI 2022 Oral session)
Geonmo Gu*, Byungsoo Ko*, SeoungHyun Go, Sung-Hyun Lee, Jingeun Lee, Minchul Shin (* Authors contributed equally.)
First figure: Comparison of M-LSD and existing LSD methods on GPU. Second figure: Inference speed and memory usage on mobile devices.
We present a real-time and light-weight line segment detector for resource-constrained environments named Mobile LSD (M-LSD). M-LSD exploits extremely efficient LSD architecture and novel training schemes, including SoL augmentation and geometric learning scheme. Our model can run in real-time on GPU, CPU, and even on mobile devices.
效果
效果1
效果2
效果3
效果4
模型信息
Inputs
-------------------------
name:input_image_with_alpha:0
tensor:Float[1, 512, 512, 4]
---------------------------------------------------------------
Outputs
-------------------------
name:Identity
tensor:Int32[1, 200, 2]
name:Identity_1
tensor:Float[1, 200]
name:Identity_2
tensor:Float[1, 256, 256, 4]
---------------------------------------------------------------
项目
VS2022
.net framework 4.8
OpenCvSharp 4.8
Microsoft.ML.OnnxRuntime 1.16.2
代码
using Microsoft.ML.OnnxRuntime.Tensors; using Microsoft.ML.OnnxRuntime; using OpenCvSharp; using System; using System.Collections.Generic; using System.Windows.Forms; using System.Linq; using System.Drawing; namespace Onnx_Demo { public partial class frmMain : Form { public frmMain() { InitializeComponent(); } string fileFilter = "*.*|*.bmp;*.jpg;*.jpeg;*.tiff;*.tiff;*.png"; string image_path = ""; DateTime dt1 = DateTime.Now; DateTime dt2 = DateTime.Now; int inpWidth; int inpHeight; Mat image; string model_path = ""; SessionOptions options; InferenceSession onnx_session; Tensor<float> input_tensor; Tensor<float> mask_tensor; List<NamedOnnxValue> input_ontainer; IDisposableReadOnlyCollection<DisposableNamedOnnxValue> result_infer; DisposableNamedOnnxValue[] results_onnxvalue; float conf_threshold = 0.5f; float dist_threshold = 20.0f; private void button1_Click(object sender, EventArgs e) { OpenFileDialog ofd = new OpenFileDialog(); ofd.Filter = fileFilter; if (ofd.ShowDialog() != DialogResult.OK) return; pictureBox1.Image = null; pictureBox2.Image = null; textBox1.Text = ""; image_path = ofd.FileName; pictureBox1.Image = new System.Drawing.Bitmap(image_path); image = new Mat(image_path); } private void Form1_Load(object sender, EventArgs e) { // 创建输入容器 input_ontainer = new List<NamedOnnxValue>(); // 创建输出会话 options = new SessionOptions(); options.LogSeverityLevel = OrtLoggingLevel.ORT_LOGGING_LEVEL_INFO; options.AppendExecutionProvider_CPU(0);// 设置为CPU上运行 // 创建推理模型类,读取本地模型文件 model_path = "model/model_512x512_large.onnx"; inpWidth = 512; inpHeight = 512; onnx_session = new InferenceSession(model_path, options); // 创建输入容器 input_ontainer = new List<NamedOnnxValue>(); image_path = "test_img/4.jpg"; pictureBox1.Image = new Bitmap(image_path); } private unsafe void button2_Click(object sender, EventArgs e) { if (image_path == "") { return; } textBox1.Text = "检测中,请稍等……"; pictureBox2.Image = null; System.Windows.Forms.Application.DoEvents(); image = new Mat(image_path); Mat resize_image = new Mat(); Cv2.Resize(image, resize_image, new OpenCvSharp.Size(512, 512)); float h_ratio = (float)image.Rows / 512; float w_ratio = (float)image.Cols / 512; int row = resize_image.Rows; int col = resize_image.Cols; float[] input_tensor_data = new float[1 * 4 * row * col]; int k = 0; for (int i = 0; i < row; i++) { for (int j = 0; j < col; j++) { for (int c = 0; c < 3; c++) { float pix = ((byte*)(resize_image.Ptr(i).ToPointer()))[j * 3 + c]; input_tensor_data[k] = pix; k++; } input_tensor_data[k] = 1; k++; } } input_tensor = new DenseTensor<float>(input_tensor_data, new[] { 1, 512, 512, 4 }); //将 input_tensor 放入一个输入参数的容器,并指定名称 input_ontainer.Add(NamedOnnxValue.CreateFromTensor("input_image_with_alpha:0", input_tensor)); dt1 = DateTime.Now; //运行 Inference 并获取结果 result_infer = onnx_session.Run(input_ontainer); dt2 = DateTime.Now; //将输出结果转为DisposableNamedOnnxValue数组 results_onnxvalue = result_infer.ToArray(); int[] pts = results_onnxvalue[0].AsTensor<int>().ToArray(); float[] pts_score = results_onnxvalue[1].AsTensor<float>().ToArray(); float[] vmap = results_onnxvalue[2].AsTensor<float>().ToArray(); List<List<int>> segments_list = new List<List<int>>(); int num_lines = 200; int map_h = 256; int map_w = 256; for (int i = 0; i < num_lines; i++) { int y = pts[i * 2]; int x = pts[i * 2 + 1]; float disp_x_start = vmap[0 + y * map_w * 4 + x * 4]; float disp_y_start = vmap[1 + y * map_w * 4 + x * 4]; float disp_x_end = vmap[2 + y * map_w * 4 + x * 4]; float disp_y_end = vmap[3 + y * map_w * 4 + x * 4]; float distance = (float)Math.Sqrt(Math.Pow(disp_x_start - disp_x_end, 2) + Math.Pow(disp_y_start - disp_y_end, 2)); if (pts_score[i] > conf_threshold && distance > dist_threshold) { float x_start = (x + disp_x_start) * 2 * w_ratio; float y_start = (y + disp_y_start) * 2 * h_ratio; float x_end = (x + disp_x_end) * 2 * w_ratio; float y_end = (y + disp_y_end) * 2 * h_ratio; List<int> line = new List<int>() { (int)x_start, (int)y_start, (int)x_end, (int)y_end }; segments_list.Add(line); } } Mat result_image = image.Clone(); for (int i = 0; i < segments_list.Count; i++) { Cv2.Line(result_image, new OpenCvSharp.Point(segments_list[i][0], segments_list[i][1]), new OpenCvSharp.Point(segments_list[i][2], segments_list[i][3]), new Scalar(0, 0, 255), 3); } pictureBox2.Image = new System.Drawing.Bitmap(result_image.ToMemoryStream()); textBox1.Text = "推理耗时:" + (dt2 - dt1).TotalMilliseconds + "ms"; } private void pictureBox2_DoubleClick(object sender, EventArgs e) { Common.ShowNormalImg(pictureBox2.Image); } private void pictureBox1_DoubleClick(object sender, EventArgs e) { Common.ShowNormalImg(pictureBox1.Image); } } }
其他
结合透视变换可实现图像校正,图像校正参考
以上就是C# Onnx实现轻量实时的M-LSD直线检测的详细内容,更多关于C#直线检测的资料请关注脚本之家其它相关文章!