C# OpenCvSharp 部署读光-票证检测矫正模型(cv_resnet18_card_correction)

说明

地址：https://modelscope.cn/models/iic/cv_resnet18_card_correction

票证检测矫正模型在实际生活中有着广泛的需求，例如信息抽取、图像质量判断、证件扫描、票据审计等领等场景，可以大幅提高工作效率和准确性。本次 读光团队 开源了商用票证检测矫正模型，基于海量的真实数据训练，可以从容应对多种复杂场景的票证检测矫正任务，该模型具有以下优点：

• 支持任意角度、多卡证票据等混贴场景，同时检测输入图像任意角度的多个子图区域；
• 基于海量真实数据训练，效果满足国内常见的卡证票据的检测矫正需求；
• 支持子图区域复印件判断、四方向判断，准确率高达 99%；
• 矫正效果、推理速度远高于 modelScope 同类模型，详见本文测试报告。

输入图片，基于 Resnet18-FPN 提取特征后，在 1/4 尺寸处通过三条分支分别识别出票证的中心点、偏移量（中心点到4个顶点距离）、中心点偏移量（为了得到精准的中心点），即可解码数出票证区域的四边形框；再用透视变换将票证拉平得到矫正后的票证信息；与此同时，分类分支识别出子图朝向，用于而切割的子图转正。

算法流程

测试时的主要预处理和后处理如下：

• 图像预处理：将输入图片按照比例缩放，长边 Resize 到 768，短边 Pad 到长短边相等，同时有减均值、除方差等归一化操作。
• 模型卡证区域检测：对输入图像中的卡证票据区域进行检测，并对 卡证票据的方向 和 复印件类型 进行分类；
• 后处理矫正：根据 卡证区域检测框 和 卡证票据的方向 对卡证区域进行透视变化，并转为水平方向；

输出字段定义

所有字段第一个维度的长度相等且一一对应，为图片中票证的数量。比如polygons[0]、scores[0]、labels[0]、layout[0]、output_imgs[0]表示第一个子图的坐标、置信度、方向、是否复印件、拉平后的子图。

效果

模型信息

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

Inputs
-------------------------
name：input
tensor：Float[1, 3, 768, 768]
---------------------------------------------------------------

Outputs
-------------------------
name：output
tensor：Float[1, 1, 192, 192]
name：286
tensor：Float[1, 4, 192, 192]
name：289
tensor：Float[1, 2, 192, 192]
name：292
tensor：Float[1, 8, 192, 192]
name：295
tensor：Float[1, 2, 192, 192]
---------------------------------------------------------------

项目

代码

using OpenCvSharp;
using System;
using System.Collections.ObjectModel;
using System.Diagnostics;
using System.Drawing;
using System.Runtime.InteropServices;
using System.Text;
using System.Windows.Forms;


namespace C__OpenCvSharp_DNN_卡证检测矫正
{
 public partial class Form1 : Form
 {
 public Form1()
 {
 InitializeComponent();
 }

 Stopwatch stopwatch = new Stopwatch();
 Mat image;
 string image_path;
 string startupPath;
 string model_path;
 string fileFilter = "*.*|*.bmp;*.jpg;*.jpeg;*.tiff;*.tiff;*.png";
 const string DllName = "CardCorrectionSharp.dll";
 IntPtr engine;
 /*
 //初始化
 extern "C" _declspec(dllexport) int __cdecl init(void** engine, char* model_path, char* msg);

 //校准
 extern "C" _declspec(dllexport) int __cdecl correction(void* engine, Mat* image, char* msg, int* out_imgs_size, Mat* out_img1, Mat* out_img2, Mat* out_img3);

 //释放
 extern "C" _declspec(dllexport) void __cdecl destroy(void* engine);

 */

 [DllImport(DllName, EntryPoint = "init", CallingConvention = CallingConvention.Cdecl)]
 internal extern static int init(ref IntPtr engine, string model_path, StringBuilder msg);

 [DllImport(DllName, EntryPoint = "correction", CallingConvention = CallingConvention.Cdecl)]
 internal extern static int correction(IntPtr engine, IntPtr image, StringBuilder msg, ref int out_imgs_size, IntPtr out_img1, IntPtr out_img2, IntPtr out_img3);

 [DllImport(DllName, EntryPoint = "destroy", CallingConvention = CallingConvention.Cdecl)]
 internal extern static int destroy(IntPtr engine);

 private void Form1_Load(object sender, EventArgs e)
 {
 startupPath = Application.StartupPath;

 model_path = startupPath + "\\model\\cv_resnet18_card_correction.onnx";

 StringBuilder msg = new StringBuilder(512);

 int res = init(ref engine, model_path, msg);
 if (res == -1)
 {
 MessageBox.Show(msg.ToString());
 return;
 }
 else
 {
 Console.WriteLine(msg.ToString());
 }
 image_path = startupPath + "\\test_img\\1.jpg";
 pictureBox1.Image = new Bitmap(image_path);
 image = new Mat(image_path);
 }

 private void button1_Click(object sender, EventArgs e)
 {
 OpenFileDialog ofd = new OpenFileDialog();
 ofd.Filter = fileFilter;
 if (ofd.ShowDialog() != DialogResult.OK) return;

 pictureBox1.Image = ;
 pictureBox2.Image = ;
 textBox1.Text = "";

 image_path = ofd.FileName;
 pictureBox1.Image = new Bitmap(image_path);
 image = new Mat(image_path);
 }

 private void Form1_FormClosing(object sender, FormClosingEventArgs e)
 {
 destroy(engine);
 }

 Mat out_img1;
 Mat out_img2;
 Mat out_img3;

 private void button2_Click(object sender, EventArgs e)
 {
 if (image_path == "")
 {
 return;
 }

 button2.Enabled = false;

 Application.DoEvents();

 Cv2.DestroyAllWindows();
 if (image!=) image.Dispose();
 if (out_img1 != ) out_img1.Dispose();
 if (out_img2 != ) out_img2.Dispose();
 if (out_img3 != ) out_img3.Dispose();
 if (pictureBox1.Image != ) pictureBox1.Image.Dispose();

 StringBuilder msg = new StringBuilder(512);
 int out_imgs_size = 0;
 image = new Mat(image_path);
 out_img1 = new Mat();
 out_img2 = new Mat();
 out_img3 = new Mat();

 stopwatch.Restart();

 int res = correction(engine, image.CvPtr, msg, ref out_imgs_size, out_img1.CvPtr, out_img2.CvPtr, out_img3.CvPtr);
 if (res == 0)
 {
 stopwatch.Stop();
 double costTime = stopwatch.Elapsed.TotalMilliseconds;
 if (out_imgs_size >= 1)
 {
 pictureBox2.Image = new Bitmap(out_img1.ToMemoryStream());
 }

 if (out_imgs_size >= 2)
 {
 Cv2.ImShow("2", out_img2);
 }

 if (out_imgs_size >= 3)
 {
 Cv2.ImShow("3", out_img3);
 }

 textBox1.Text = $"耗时:{costTime:F2}ms";
 }
 else
 {
 textBox1.Text = "识别失败";
 }
 button2.Enabled = true;
 }
 }
}

参考

https://github.com/hpc203/cv_resnet18_card_correction-opencv-dnn