[C#/WINFORM] 텐서플로우를 사용해 물체 인식하기

TestProject.zip

다운로드

TestProject.z01

다운로드

TestProject.z02

다운로드

TestProject.z03

다운로드

TestProject.z04

다운로드

TestProject.z05

다운로드

TestProject.z06

다운로드

TestProject.z07

다운로드

TestProject.z08

다운로드

TestProject.z09

다운로드

TestProject.z10

다운로드

TestProject.z11

다운로드

TestProject.z12

다운로드

TestProject.z13

다운로드

TestProject.z14

다운로드

▶ MainForm.cs

using System;
using System.Drawing;
using System.Drawing.Imaging;
using System.IO;
using System.Windows.Forms;

using AForge.Video;
using AForge.Video.DirectShow;

using TensorFlow;

namespace TestProject
{
    /// <summary>
    /// 메인 폼
    /// </summary>
    public partial class MainForm : Form
    {
        //////////////////////////////////////////////////////////////////////////////////////////////////// Field
        ////////////////////////////////////////////////////////////////////////////////////////// Private

        #region Field

        /// <summary>
        /// 디스플레이 레이블 그래픽스
        /// </summary>
        private Graphics displayLabelGraphics = null;

        /// <summary>
        /// 필터 정보 컬렉션
        /// </summary>
        private FilterInfoCollection filterInfoCollection = null;

        /// <summary>
        /// 비디오 캡처 장치
        /// </summary>
        private VideoCaptureDevice videoCaptureDevice = null;


        /// <summary>
        /// 모델 배열
        /// </summary>
        private byte[] modelArray = null;

        /// <summary>
        /// 레이블 배열
        /// </summary>
        private string[] labelArray = null;


        /// <summary>
        /// 텐서플로우 그래프
        /// </summary>
        private TFGraph graph = null;

        /// <summary>
        /// 텐서플로우 세션
        /// </summary>
        private TFSession session = null;

        #endregion

        //////////////////////////////////////////////////////////////////////////////////////////////////// Constructor
        ////////////////////////////////////////////////////////////////////////////////////////// Public

        #region MainForm

        /// <summary>
        /// 생성자
        /// </summary>
        public MainForm()
        {
            InitializeComponent();

            #region 이벤트를 설정한다.

            Load                   += Form_Load;
            FormClosing            += Form_FormClosing;
            this.startButton.Click += startButton_Click;

            #endregion
        }

        #endregion

        //////////////////////////////////////////////////////////////////////////////////////////////////// Method
        ////////////////////////////////////////////////////////////////////////////////////////// Private
        //////////////////////////////////////////////////////////////////////////////// Event

        #region 폼 로드시 처리하기 - Form_Load(sender, e)

        /// <summary>
        /// 폼 로드시 처리하기
        /// </summary>
        /// <param name="sender">이벤트 발생자</param>
        /// <param name="e">이벤트 인자</param>
        private void Form_Load(object sender, EventArgs e)
        {
            this.displayLabelGraphics = this.displayLabel.CreateGraphics();

            #region 카메라 장치 콤보 박스를 설정한다.

            this.filterInfoCollection = new FilterInfoCollection(FilterCategory.VideoInputDevice);

            for(int i = 0; i < this.filterInfoCollection.Count; i++)
            {
                this.cameraDeviceComboBox.Items.Add(this.filterInfoCollection[i].Name);
            }

            if(this.cameraDeviceComboBox.Items.Count > 0)
            {
                this.cameraDeviceComboBox.SelectedIndex = 0;
            }

            #endregion
            #region 시작 버튼을 설정한다.

            this.startButton.Enabled = (this.cameraDeviceComboBox.Items.Count > 0);

            #endregion

            #region 모델 배열을 설정한다.

            this.modelArray = File.ReadAllBytes("DATA\\frozen_process_no_filter_tiny.pb");

            #endregion
            #region 레이블 배열을 설정한다.

            this.labelArray = File.ReadAllLines("DATA\\yolo_labels.txt");

            #endregion
            #region 텐서플로우 그래프를 설정한다.

            this.graph = new TFGraph();

            this.graph.Import(this.modelArray, "");

            #endregion
            #region 텐서플로우 세션을 설정한다.

            this.session = new TFSession(graph);

            #endregion

            Bitmap bitmap = new Bitmap(100, 100, PixelFormat.Format24bppRgb);

            Predict(bitmap);
        }

        #endregion
        #region 폼 닫을 경우 처리하기 - Form_FormClosing(sender, e)

        /// <summary>
        /// 폼 닫을 경우 처리하기
        /// </summary>
        /// <param name="sender">이벤트 발생자</param>
        /// <param name="e">이벤트 인자</param>
        private void Form_FormClosing(object sender, FormClosingEventArgs e)
        {
            StopCameraCaoture();
        }

        #endregion
        #region 시작 버튼 클릭시 처리하기 - startButton_Click(sender, e)

        /// <summary>
        /// 시작 버튼 클릭시 처리하기
        /// </summary>
        /// <param name="sender">이벤트 발생자</param>
        /// <param name="e">이벤트 인자</param>
        private void startButton_Click(object sender, EventArgs e)
        {
            if(this.startButton.Text == "시작")
            {
                this.startButton.Text = "중지";

                StartCameraCapture();
            }
            else
            {
                StopCameraCaoture();

                this.startButton.Text = "시작";
            }
        }

        #endregion
        #region 비디오 캡처 장치 신규 프레임 처리하기 - videoCaptureDevice_NewFrame(sender, e)

        /// <summary>
        /// 비디오 캡처 장치 신규 프레임 처리하기
        /// </summary>
        /// <param name="sender">이벤트 발생자</param>
        /// <param name="e">이벤트 인자</param>
        private void videoCaptureDevice_NewFrame(object sender, NewFrameEventArgs e)
        {
            Bitmap bitmap = e.Frame;

            TFTensor[] outputTensorArray = Predict(bitmap);

            float[,] rectangleArray   = (float[,])outputTensorArray[0].GetValue();
            float[]  probabilityArray = (float[] )outputTensorArray[1].GetValue();
            long[]   labelIndexArray  = (long[]  )outputTensorArray[2].GetValue();

            using(Graphics graphics = Graphics.FromImage(bitmap))
            {
                for(int i = 0; i < probabilityArray.Length; i++)
                {
                    if(probabilityArray[i] > 0.2)
                    {
                        int rectangleLeft   = (int)(rectangleArray[i, 0] * (float)bitmap.Width );
                        int rectangleTop    = (int)(rectangleArray[i, 1] * (float)bitmap.Height);
                        int rectangleWidth  = (int)(rectangleArray[i, 2] * (float)bitmap.Width );
                        int rectangleHeight = (int)(rectangleArray[i, 3] * (float)bitmap.Height);

                        graphics.DrawRectangle
                        (
                            new Pen(Color.Red, 2),
                            rectangleLeft - rectangleWidth  / 2,
                            rectangleTop  - rectangleHeight / 2,
                            rectangleWidth,
                            rectangleHeight
                        );

                        string label = this.labelArray[labelIndexArray[i]] + " " + probabilityArray[i].ToString();

                        graphics.DrawString
                        (
                            label,
                            Font,
                            Brushes.Red,
                            rectangleLeft - rectangleWidth  / 2,
                            rectangleTop  - rectangleHeight / 2
                        );
                    }
                }
            }

            this.displayLabelGraphics.DrawImage(bitmap, 0, 0, this.displayLabel.Width, this.displayLabel.Height);
        }

        #endregion

        //////////////////////////////////////////////////////////////////////////////// Function

        #region 카메라 캡처 시작하기 - StartCameraCapture()

        /// <summary>
        /// 카메라 캡처 시작하기
        /// </summary>
        private void StartCameraCapture()
        {
            if(this.videoCaptureDevice == null || !this.videoCaptureDevice.IsRunning)
            {
                FilterInfo filterInfo = this.filterInfoCollection[this.cameraDeviceComboBox.SelectedIndex];

                this.videoCaptureDevice = new VideoCaptureDevice(filterInfo.MonikerString);

                this.videoCaptureDevice.NewFrame += videoCaptureDevice_NewFrame;

                this.videoCaptureDevice.Start();
            }
        }

        #endregion
        #region 카메라 캡처 중단하기 - StopCameraCaoture()

        /// <summary>
        /// 카메라 캡처 중단하기
        /// </summary>
        private void StopCameraCaoture()
        {
            if(this.videoCaptureDevice != null && this.videoCaptureDevice.IsRunning)
            {
                this.videoCaptureDevice.NewFrame -= videoCaptureDevice_NewFrame;

                this.videoCaptureDevice.SignalToStop();

                this.videoCaptureDevice.WaitForStop();

                this.videoCaptureDevice = null;
            }
        }

        #endregion

        #region 텐서 구하기 - GetTensor(sourceArray)

        /// <summary>
        /// 텐서 구하기
        /// </summary>
        /// <param name="sourceArray">소스 배열</param>
        /// <returns>텐서</returns>
        private TFTensor GetTensor(byte[] sourceArray)
        {
            TFGraph  graph  = new TFGraph();
            TFOutput input  = graph.Placeholder(TFDataType.String);
            TFOutput output = graph.Cast(graph.DecodeJpeg(contents : input, channels : 3), DstT : TFDataType.Float);
            TFTensor tensor = TFTensor.CreateString(sourceArray);

            using(TFSession session = new TFSession(graph))
            {
                TFTensor[] normalizedTensorArray = session.Run
                (
                    inputs      : new[] { input  },
                    outputs     : new[] { output },
                    inputValues : new[] { tensor }
                );

                return normalizedTensorArray[0];
            }
        }

        #endregion
        #region 예측하기 - Predict(bitmap)

        /// <summary>
        /// 예측하기
        /// </summary>
        /// <param name="bitmap">비트맵</param>
        /// <returns>결과 텐서 배열</returns>
        private TFTensor[] Predict(Bitmap bitmap)
        {
            MemoryStream memoryStream = new MemoryStream();

            bitmap.Save(memoryStream, ImageFormat.Jpeg);

            TFTensor tensor = GetTensor(memoryStream.GetBuffer());

            TFSession.Runner runner = this.session.GetRunner();

            runner.AddInput(this.graph["input"][0], tensor).Fetch("boxes", "classes_prob", "classes_arg");

            TFTensor[] outputTensorArray = runner.Run();

            return outputTensorArray;
        }

        #endregion
    }
}