Unity & EasyDL 图像分割 - 识别图像中主体及其位置

2022-08-29 16:36:54 浏览数 (1)

EasyGL图像分割介绍:

创建应用:

1.进入百度AI开放平台打开控制台:

2.在左上角打开产品服务列表,找到EasyDL零门槛AI开放平台:

3.打开EasyGL图像:

4.在公有云部署-应用列表中创建一个应用:

5.创建完成后获取到AppID、API Key、Secret Key:

创建模型:

1.进入EasyGL图像分割:

2.创建模型:

3.创建数据集:

4.数据导入:

上传图片,图片的数量尽量多些

导入完成后查看并标注:

框选目标所在范围:

添加标签并为框选的目标设置标签:

设置完成后保存当前标注:

5.训练模型:(开始训练后需要等待一定时间)

6.发布模型:

发布完成后,拿到接口地址,来到Unity中,根据接口响应字段说明定义相应数据结构:

代码语言:javascript复制
using System;

[Serializable]
public class ImageSegmentationResponse
{
    /// <summary>
    /// 唯一的log id 用于问题定位
    /// </summary>
    public int log_id;
    /// <summary>
    /// 标签数组结果
    /// </summary>
    public ImageSegmentationResult[] results;
}
[Serializable]
public class ImageSegmentationResult
{
    /// <summary>
    /// 标签名称
    /// </summary>
    public string name;
    /// <summary>
    /// 置信度
    /// </summary>
    public string score;
    /// <summary>
    /// 位置
    /// </summary>
    public Location location;
    /// <summary>
    /// 基于游程编码的字符串,编码内容为和原图宽高相同的布尔数组
    /// 若数组值为0,代表原图此位置像素点不属于检测目标,若为1,代表原图此位置像素点属于检测目标
    /// </summary>
    public bool[] mask;
}
[Serializable]
public class Location
{
    /// <summary>
    /// 目标定位位置的长方形左上顶点的水平坐标
    /// </summary>
    public int left;
    /// <summary>
    /// 目标定位位置的长方形左上顶点的垂直坐标
    /// </summary>
    public int top;
    /// <summary>
    /// 目标定位位置的长方形的宽度
    /// </summary>
    public int width;
    /// <summary>
    /// 目标定位位置的长方形的高度
    /// </summary>
    public int height;
}

在任意一个模块下载C#SDK,例如在图像识别中下载,它是包含EasyDL的API内容的:

有了SDK后,放入Unity中的Plugins文件夹中,封装调用函数,只需要将检测图片的字节数据作为参数,其中appID、apiKey、secretKey是在上面创建应用时获取到的,url是发布模型时获取到的:

代码语言:javascript复制
using System;
using UnityEngine;
/// <summary>
/// 图像分割
/// </summary>
public class ImageSegmentation
{
    private const string appID = "";
    private const string apiKey = "";
    private const string secretKey = "";
    private const string url = "";
    public static ImageSegmentationResult[] SendRequest(byte[] bytes)
    {
        var client = new Baidu.Aip.EasyDL.EasyDL(appID, apiKey, secretKey);
        try
        {
            var response = client.requestImage(url, bytes);
            Debug.Log(response.ToString());
            ImageSegmentationResponse r = JsonUtility.FromJson<ImageSegmentationResponse>(response.ToString());
            return r.results;
        }
        catch (Exception error)
        {
            Debug.LogError(error);
        }
        return null;
    }
}

测试图片:

测试代码:

代码语言:javascript复制
using System.IO;
using UnityEngine;

public class Example : MonoBehaviour
{
    private void Start()
    {
        ImageSegmentation.SendRequest(File.ReadAllBytes(Application.dataPath   "/1.jpg"));
    }
}

返回结果:

拿到了定位数据后,接下来将其区域绘制出来, 响应说明中解释(left,top)构成左上顶点,但是从返回值来看top为16,减去一个高度312的话,左下顶点的坐标已经是负数,这里姑且猜想它构成的是左下顶点:

首先创建一个Image来放置我们的测试图片,Canvas、Image大小也设为测试图片的大小640 * 359:

以下是测试脚本,将其挂载于Image测试:

代码语言:javascript复制
using System.IO;
using UnityEngine;

public class Example : MonoBehaviour
{
    private void Start()
    {
        var results = ImageSegmentation.SendRequest(File.ReadAllBytes(Application.dataPath   "/测试.jpg"));

        for (int i = 0; i < results.Length; i  )
        {
            var location = results[i].location;

            LineRenderer line = new GameObject("LineRenderer").AddComponent<LineRenderer>();
            line.positionCount = 4;
            line.loop = true;

            Vector2 leftTop = new Vector2(location.left, location.top);
            Vector2 rightTop = new Vector2(location.left   location.width, location.top);
            Vector2 leftBottom = new Vector2(location.left, location.top   location.height);
            Vector2 rightBottom = new Vector2(location.left   location.width, location.top   location.height);

            RectTransformUtility.ScreenPointToWorldPointInRectangle(transform as RectTransform, leftTop, Camera.main, out Vector3 point1);
            RectTransformUtility.ScreenPointToWorldPointInRectangle(transform as RectTransform, rightTop, Camera.main, out Vector3 point2);
            RectTransformUtility.ScreenPointToWorldPointInRectangle(transform as RectTransform, rightBottom, Camera.main, out Vector3 point3);
            RectTransformUtility.ScreenPointToWorldPointInRectangle(transform as RectTransform, leftBottom, Camera.main, out Vector3 point4);

            line.SetPosition(0, point1);
            line.SetPosition(1, point2);
            line.SetPosition(2, point3);
            line.SetPosition(3, point4);
        }
    }
}

emmm... 区域大概准确吧,可能测试的模型数据集足够丰富的话检测会更精确。

0 人点赞