C# CvDnn部署CoupledTPS实现旋转图像矫正

寒冰屋

已于 2024-07-10 14:59:50 修改

阅读量2

点赞数

分类专栏： CSharp.NET 人工智能文章标签： c# CvDnn CoupledTPS

于 2024-07-10 14:55:43 首次发布

原文链接：https://mp.weixin.qq.com/s?__biz=MzAwNTMxMzg1MA==&mid=2654099804&idx=8&sn=1e4b9b0389205512b357a86186a52cce&chksm=815f2276cfe081aa48cded74636bb29747a1d1cac9ae833fa8d6c125a8bd8c3fa7a2ff62b98e&scene=126&sessionid=1720590601#rd

版权

CSharp.NET 同时被 2 个专栏收录

1136 篇文章 52 订阅

订阅专栏

人工智能

513 篇文章 54 订阅

订阅专栏

介绍

TPAMI2024 - Semi-Supervised Coupled Thin-Plate Spline Model for Rotation Correction and Beyond

github地址：

https://github.com/nie-lang/CoupledTPS

代码实现参考：

https://github.com/hpc203/CoupledTPS-opencv-dnn

效果

模型信息

feature_extractor.onnx

Model Properties
-------------------------
---------------------------------------------------------------

Inputs
-------------------------
name：input
tensor：Float[1, 3, 384, 512]
---------------------------------------------------------------

Outputs
-------------------------
name：feature
tensor：Float[1, 256, 24, 32]
---------------------------------------------------------------


regressnet.onnx


Model Properties
-------------------------
---------------------------------------------------------------

Inputs
-------------------------
name：feature
tensor：Float[1, 256, 24, 32]
---------------------------------------------------------------

Outputs
-------------------------
name：mesh_motion
tensor：Float[1, 7, 9, 2]
---------------------------------------------------------------

项目

VS2022

.net framework 4.8

OpenCvSharp 4.8

代码

Form1.cs

using OpenCvSharp;
using System;
using System.Drawing;
using System.Drawing.Imaging;
using System.Windows.Forms;
 
namespace Onnx_Demo
{
    public partial class Form1 : Form
    {
        public Form1()
        {
            InitializeComponent();
        }
 
        string fileFilter = "*.*|*.bmp;*.jpg;*.jpeg;*.tiff;*.tiff;*.png";
        string image_path = "";
        DateTime dt1 = DateTime.Now;
        DateTime dt2 = DateTime.Now;
        Mat image;
 
        CoupledTPS_RotationNet rotationNet;
        int iter_num = 3;
 
        private void button1_Click(object sender, EventArgs e)
        {
            OpenFileDialog ofd = new OpenFileDialog();
            ofd.InitialDirectory =Application.StartupPath+"\\test_img\\";
            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;
        }
 
        private void button2_Click(object sender, EventArgs e)
        {
            if (image_path == "")
            {
                return;
            }
            button2.Enabled = false;
            pictureBox2.Image = null;
            textBox1.Text = "";
            Application.DoEvents();
            //读图片
            image = new Mat(image_path);
            dt1 = DateTime.Now;
            Mat result_image = rotationNet.detect(image, iter_num);
            dt2 = DateTime.Now;
            Cv2.CvtColor(result_image, result_image, ColorConversionCodes.BGR2RGB);
            pictureBox2.Image = new Bitmap(result_image.ToMemoryStream());
            textBox1.Text = "推理耗时:" + (dt2 - dt1).TotalMilliseconds + "ms";
            button2.Enabled = true;
        }
 
        private void Form1_Load(object sender, EventArgs e)
        {
            rotationNet = new CoupledTPS_RotationNet("model/feature_extractor.onnx", "model/regressnet.onnx");
            image_path = "test_img/00150_-8.4.jpg";
            pictureBox1.Image = new Bitmap(image_path);
            image = new Mat(image_path);
        }
 
        private void pictureBox1_DoubleClick(object sender, EventArgs e)
        {
            Common.ShowNormalImg(pictureBox1.Image);
        }
 
        private void pictureBox2_DoubleClick(object sender, EventArgs e)
        {
            Common.ShowNormalImg(pictureBox2.Image);
        }
 
        SaveFileDialog sdf = new SaveFileDialog();
        private void button3_Click(object sender, EventArgs e)
        {
            if (pictureBox2.Image == null)
            {
                return;
            }
            Bitmap output = new Bitmap(pictureBox2.Image);
            sdf.Title = "保存";
            sdf.Filter = "Images (*.jpg)|*.jpg|Images (*.png)|*.png|Images (*.bmp)|*.bmp|Images (*.emf)|*.emf|Images (*.exif)|*.exif|Images (*.gif)|*.gif|Images (*.ico)|*.ico|Images (*.tiff)|*.tiff|Images (*.wmf)|*.wmf";
            if (sdf.ShowDialog() == DialogResult.OK)
            {
                switch (sdf.FilterIndex)
                {
                    case 1:
                        {
                            output.Save(sdf.FileName, ImageFormat.Jpeg);
                            break;
                        }
                    case 2:
                        {
                            output.Save(sdf.FileName, ImageFormat.Png);
                            break;
                        }
                    case 3:
                        {
                            output.Save(sdf.FileName, ImageFormat.Bmp);
                            break;
                        }
                    case 4:
                        {
                            output.Save(sdf.FileName, ImageFormat.Emf);
                            break;
                        }
                    case 5:
                        {
                            output.Save(sdf.FileName, ImageFormat.Exif);
                            break;
                        }
                    case 6:
                        {
                            output.Save(sdf.FileName, ImageFormat.Gif);
                            break;
                        }
                    case 7:
                        {
                            output.Save(sdf.FileName, ImageFormat.Icon);
                            break;
                        }
 
                    case 8:
                        {
                            output.Save(sdf.FileName, ImageFormat.Tiff);
                            break;
                        }
                    case 9:
                        {
                            output.Save(sdf.FileName, ImageFormat.Wmf);
                            break;
                        }
                }
                MessageBox.Show("保存成功，位置：" + sdf.FileName);
            }
        }
    }
}

CoupledTPS_RotationNet.cs

using OpenCvSharp;
using OpenCvSharp.Dnn;
using System.Collections.Generic;
using System.Linq;
 
namespace Onnx_Demo
{
    public class CoupledTPS_RotationNet
    {
 
        int input_height = 384;
        int input_width = 512;
        int grid_h = 6;
        int grid_w = 8;
        Mat grid = new Mat();
        Mat W_inv = new Mat();
 
        Net feature_extractor;
        Net regressNet;
 
        public CoupledTPS_RotationNet(string modelpatha, string modelpathb)
        {
            feature_extractor = CvDnn.ReadNet(modelpatha);
            regressNet = CvDnn.ReadNet(modelpathb);
            tps2flow.get_norm_rigid_mesh_inv_grid(ref grid, ref W_inv, input_height, input_width, grid_h, grid_w);
        }
 
        unsafe public Mat detect(Mat srcimg, int iter_num)
        {
            Mat img = new Mat();
            Cv2.Resize(srcimg, img, new Size(input_width, input_height));
            img.ConvertTo(img, MatType.CV_32FC3, 1.0 / 127.5d, -1.0d);
 
            Mat input_tensor = CvDnn.BlobFromImage(img);
 
            feature_extractor.SetInput(input_tensor);
 
            Mat[] feature_oris = new Mat[1] { new Mat() };
            string[] outBlobNames = feature_extractor.GetUnconnectedOutLayersNames().ToArray();
            feature_extractor.Forward(feature_oris, outBlobNames);
            Mat feature = feature_oris[0].Clone();
 
            int[] shape = { 1, 2, input_height, input_width };
            Mat flow = Mat.Zeros(MatType.CV_32FC1, shape);
 
            List<Mat> flow_list = new List<Mat>();
            for (int i = 0; i < iter_num; i++)
            {
                regressNet.SetInput(feature);
                Mat[] mesh_motions = new Mat[1] { new Mat() };
                regressNet.Forward(mesh_motions, regressNet.GetUnconnectedOutLayersNames().ToArray());
 
                float* offset = (float*)mesh_motions[0].Data;
 
                Mat tp = new Mat();
 
                tps2flow.get_ori_rigid_mesh_tp(ref tp, offset, input_height, input_width, grid_h, grid_w);
 
                Mat T = W_inv * tp;   //_solve_system
                T = T.T();    //舍弃batchsize
 
                Mat T_g = T * grid;
 
                Mat delta_flow = new Mat();
 
                tps2flow._transform(T_g, grid, input_height, input_width, ref delta_flow);
 
                if (i == 0)
                {
                    flow += delta_flow;
 
                }
                else
                {
                    Mat warped_flow = new Mat();
                    grid_sample.warp_with_flow(flow, delta_flow, ref warped_flow);
 
                    flow = delta_flow + warped_flow;
                }
                flow_list.Add(flow.Clone());
 
                if (i < (iter_num - 1))
                {
                    int fea_h = feature.Size(2);
                    int fea_w = feature.Size(3);
                    float scale_h = (float)fea_h / flow.Size(2);
                    float scale_w = (float)fea_w / flow.Size(3);
 
                    Mat down_flow = new Mat();
                    upsample.UpSamplingBilinear(flow, ref down_flow, fea_h, fea_w, true, scale_h, scale_w);
 
                    for (int h = 0; h < fea_h; h++)
                    {
                        for (int w = 0; w < fea_w; w++)
                        {
                            float* p_w = (float*)down_flow.Ptr(0, 0, h);
                            float temp_w = p_w[w];
                            temp_w = temp_w * scale_w;
                            p_w[w] = temp_w;
 
                            float* p_h = (float*)down_flow.Ptr(0, 1, h);
                            float temp_h = p_h[w];
                            temp_h = temp_h * scale_h;
                            p_h[w] = temp_h;
 
                        }
                    }
                    feature.Release();
                    feature = new Mat();
                    grid_sample.warp_with_flow(feature_oris[0], down_flow, ref feature);
                }
            }
            Mat correction_final = new Mat();
 
            grid_sample.warp_with_flow(input_tensor, flow_list[iter_num - 1], ref correction_final);
 
            Mat correction_img = grid_sample.convert4dtoimage(correction_final);
 
            return correction_img;
 
        }
    }
}

引入地址