主体分割技术，提升图像信息提取能力

2025-01-10
贵州
本文字数：2157 字
阅读完需：约 7 分钟

在智能设备普及和 AI 技术进步的推动下，用户对线上互动的质量、个性化以及沉浸式体验的追求日益增强。例如，对于热衷于图片编辑或视频制作的用户来说，他们需要一种快速而简便的方法来将特定主体从背景中分离出来。HarmonyOS SDK 基础视觉服务（Core Vision Kit）提供主体分割能力，可以检测出图片中区别于背景的前景物体或区域（即"显著主体"），并将其从背景中分离出来，适用于需要识别和提取图像主要信息的场景，广泛使用于前景目标检测和前景主体分离的场景。

适用场景

主体贴纸：从图片中提取显著性的主体，去掉背景。

背景替换：替换并提取出主体对象的背景。

显著性检测：快速定位图片中显著性区域。

辅助图片编辑：例如单独对主体进行美化处理。

功能演示

本文仅展示通用三个入口：复制、分享、识图搜索，其他功能开发者可以自行定制。

开发步骤

1.引用相关类添加至工程。

import { subjectSegmentation } from '@kit.CoreVisionKit';

复制代码

2.准备预处理的图片资源，将图片转换为 PixelMap，并添加初始化和释放方法。

async aboutToAppear(): Promise<void> {  const initResult = await subjectSegmentation.init();  hilog.info(0x0000, 'subjectSegmentationSample', `Subject segmentation initialization result:${initResult}`);}
async aboutToDisappear(): Promise<void> {  await subjectSegmentation.release();  hilog.info(0x0000, 'subjectSegmentationSample', 'Subject segmentation released successfully');}
private async selectImage() {  let uri = await this.openPhoto()  if (uri === undefined) {    hilog.error(0x0000, TAG, "uri is undefined");  }  this.loadImage(uri);}
private openPhoto(): Promise<Array<string>> {  return new Promise<Array<string>>((resolve, reject) => {    let PhotoSelectOptions = new photoAccessHelper.PhotoSelectOptions();    PhotoSelectOptions.MIMEType = photoAccessHelper.PhotoViewMIMETypes.IMAGE_TYPE;    PhotoSelectOptions.maxSelectNumber = 1;    let photoPicker: photoAccessHelper.PhotoViewPicker = new photoAccessHelper.PhotoViewPicker();    hilog.info(0x0000, TAG, 'PhotoViewPicker.select successfully, PhotoSelectResult uri: ');    photoPicker.select(PhotoSelectOptions).then((PhotoSelectResult) => {      hilog.info(0x0000, TAG, `PhotoViewPicker.select successfully, PhotoSelectResult uri: ${PhotoSelectResult.photoUris}`);      resolve(PhotoSelectResult.photoUris)    }).catch((err: BusinessError) => {      hilog.error(0x0000, TAG, `PhotoViewPicker.select failed with errCode: ${err.code}, errMessage: ${err.message}`);      reject();    });  })}
private loadImage(names: string[]) {  setTimeout(async () => {    let imageSource: image.ImageSource | undefined = undefined    let fileSource = await fileIo.open(names[0], fileIo.OpenMode.READ_ONLY)    imageSource = image.createImageSource(fileSource.fd)    this.chooseImage = await imageSource.createPixelMap()    hilog.info(0x0000, TAG, `this.chooseImage===${this.chooseImage}`);  }, 100  )}

复制代码

3.实例化待分割的入参项 VisionInfo，并传入待检测图片的 PixelMap。

let visionInfo: subjectSegmentation.VisionInfo = {  pixelMap: this.chooseImage,};

复制代码

4.配置通用文本识别的配置项 SegmentationConfig，包括最大分割主体个数、是否输出每个主体的分割信息，以及是否输出分割后的前景图。

let config: subjectSegmentation.SegmentationConfig = {  maxCount: parseInt(this.maxNum),  enableSubjectDetails: true,  enableSubjectForegroundImage: true,};

复制代码

5.调用 imageSegmentation 的 ai.vision.doSegmentation 接口，实现主体分割。

let data: subjectSegmentation.SegmentationResult = await subjectSegmentation.doSegmentation(visionInfo, config);let outputString = `Subject count: ${data.subjectCount}\n`;outputString += `Max subject count: ${config.maxCount}\n`;outputString += `Enable subject details: ${config.enableSubjectDetails ? 'Yes' : 'No'}\n\n`;let segBox : subjectSegmentation.Rectangle = data.fullSubject.subjectRectangle;let segBoxString = `Full subject box:\nLeft: ${segBox.left}, Top: ${segBox.top}, Width: ${segBox.width}, Height: ${segBox.height}\n\n`;outputString += segBoxString;
if (config.enableSubjectDetails) {  outputString += 'Individual subject boxes:\n';  if (data.subjectDetails) {    for (let i = 0; i < data.subjectDetails.length; i++) {      let detailSegBox: subjectSegmentation.Rectangle = data.subjectDetails[i].subjectRectangle;      outputString += `Subject ${i + 1}:\nLeft: ${detailSegBox.left}, Top: ${detailSegBox.top}, Width: ${detailSegBox.width}, Height: ${detailSegBox.height}\n\n`;    }  }}