在 iOS 上使用机器学习套件识别数字墨水

利用机器学习套件的数字墨水识别功能,您可以识别在数以百计的数码表面手写的文字,以及对草图进行分类。

试试看

准备工作

  1. 在 Podfile 中添加以下机器学习套件库:

    pod 'GoogleMLKit/DigitalInkRecognition', '3.2.0'
    
    
  2. 安装或更新项目的 Pod 之后,请使用 Xcode 项目的 .xcworkspace 来打开项目。Xcode 版本 13.2.1 或更高版本支持机器学习套件。

您现在可以开始识别 Ink 对象中的文本。

构建 Ink 对象

构建 Ink 对象的主要方法是在触摸屏上绘制它。在 iOS 上,您可以使用 UIImageView触摸事件处理脚本来绘制屏幕上的描边,并存储描边的点以构建 Ink 对象。以下代码段对此常规模式进行了演示。请参阅快速入门应用,了解将触摸事件处理、屏幕绘制和描边数据管理分开的更完整示例。

Swift

@IBOutlet weak var mainImageView: UIImageView!
var kMillisecondsPerTimeInterval = 1000.0
var lastPoint = CGPoint.zero
private var strokes: [Stroke] = []
private var points: [StrokePoint] = []

func drawLine(from fromPoint: CGPoint, to toPoint: CGPoint) {
  UIGraphicsBeginImageContext(view.frame.size)
  guard let context = UIGraphicsGetCurrentContext() else {
    return
  }
  mainImageView.image?.draw(in: view.bounds)
  context.move(to: fromPoint)
  context.addLine(to: toPoint)
  context.setLineCap(.round)
  context.setBlendMode(.normal)
  context.setLineWidth(10.0)
  context.setStrokeColor(UIColor.white.cgColor)
  context.strokePath()
  mainImageView.image = UIGraphicsGetImageFromCurrentImageContext()
  mainImageView.alpha = 1.0
  UIGraphicsEndImageContext()
}

override func touchesBegan(_ touches: Set, with event: UIEvent?) {
  guard let touch = touches.first else {
    return
  }
  lastPoint = touch.location(in: mainImageView)
  let t = touch.timestamp
  points = [StrokePoint.init(x: Float(lastPoint.x),
                             y: Float(lastPoint.y),
                             t: Int(t * kMillisecondsPerTimeInterval))]
  drawLine(from:lastPoint, to:lastPoint)
}

override func touchesMoved(_ touches: Set, with event: UIEvent?) {
  guard let touch = touches.first else {
    return
  }
  let currentPoint = touch.location(in: mainImageView)
  let t = touch.timestamp
  points.append(StrokePoint.init(x: Float(currentPoint.x),
                                 y: Float(currentPoint.y),
                                 t: Int(t * kMillisecondsPerTimeInterval)))
  drawLine(from: lastPoint, to: currentPoint)
  lastPoint = currentPoint
}

override func touchesEnded(_ touches: Set, with event: UIEvent?) {
  guard let touch = touches.first else {
    return
  }
  let currentPoint = touch.location(in: mainImageView)
  let t = touch.timestamp
  points.append(StrokePoint.init(x: Float(currentPoint.x),
                                 y: Float(currentPoint.y),
                                 t: Int(t * kMillisecondsPerTimeInterval)))
  drawLine(from: lastPoint, to: currentPoint)
  lastPoint = currentPoint
  strokes.append(Stroke.init(points: points))
  self.points = []
  doRecognition()
}

Objective-C

// Interface
@property (weak, nonatomic) IBOutlet UIImageView *mainImageView;
@property(nonatomic) CGPoint lastPoint;
@property(nonatomic) NSMutableArray *strokes;
@property(nonatomic) NSMutableArray *points;

// Implementations
static const double kMillisecondsPerTimeInterval = 1000.0;

- (void)drawLineFrom:(CGPoint)fromPoint to:(CGPoint)toPoint {
  UIGraphicsBeginImageContext(self.mainImageView.frame.size);
  [self.mainImageView.image drawInRect:CGRectMake(0, 0, self.mainImageView.frame.size.width,
                                                  self.mainImageView.frame.size.height)];
  CGContextMoveToPoint(UIGraphicsGetCurrentContext(), fromPoint.x, fromPoint.y);
  CGContextAddLineToPoint(UIGraphicsGetCurrentContext(), toPoint.x, toPoint.y);
  CGContextSetLineCap(UIGraphicsGetCurrentContext(), kCGLineCapRound);
  CGContextSetLineWidth(UIGraphicsGetCurrentContext(), 10.0);
  CGContextSetRGBStrokeColor(UIGraphicsGetCurrentContext(), 1, 1, 1, 1);
  CGContextSetBlendMode(UIGraphicsGetCurrentContext(), kCGBlendModeNormal);
  CGContextStrokePath(UIGraphicsGetCurrentContext());
  CGContextFlush(UIGraphicsGetCurrentContext());
  self.mainImageView.image = UIGraphicsGetImageFromCurrentImageContext();
  UIGraphicsEndImageContext();
}

- (void)touchesBegan:(NSSet *)touches withEvent:(nullable UIEvent *)event {
  UITouch *touch = [touches anyObject];
  self.lastPoint = [touch locationInView:self.mainImageView];
  NSTimeInterval time = [touch timestamp];
  self.points = [NSMutableArray array];
  [self.points addObject:[[MLKStrokePoint alloc] initWithX:self.lastPoint.x
                                                         y:self.lastPoint.y
                                                         t:time * kMillisecondsPerTimeInterval]];
  [self drawLineFrom:self.lastPoint to:self.lastPoint];
}

- (void)touchesMoved:(NSSet *)touches withEvent:(nullable UIEvent *)event {
  UITouch *touch = [touches anyObject];
  CGPoint currentPoint = [touch locationInView:self.mainImageView];
  NSTimeInterval time = [touch timestamp];
  [self.points addObject:[[MLKStrokePoint alloc] initWithX:currentPoint.x
                                                         y:currentPoint.y
                                                         t:time * kMillisecondsPerTimeInterval]];
  [self drawLineFrom:self.lastPoint to:currentPoint];
  self.lastPoint = currentPoint;
}

- (void)touchesEnded:(NSSet *)touches withEvent:(nullable UIEvent *)event {
  UITouch *touch = [touches anyObject];
  CGPoint currentPoint = [touch locationInView:self.mainImageView];
  NSTimeInterval time = [touch timestamp];
  [self.points addObject:[[MLKStrokePoint alloc] initWithX:currentPoint.x
                                                         y:currentPoint.y
                                                         t:time * kMillisecondsPerTimeInterval]];
  [self drawLineFrom:self.lastPoint to:currentPoint];
  self.lastPoint = currentPoint;
  if (self.strokes == nil) {
    self.strokes = [NSMutableArray array];
  }
  [self.strokes addObject:[[MLKStroke alloc] initWithPoints:self.points]];
  self.points = nil;
  [self doRecognition];
}

请注意,该代码段包含一个将描边绘制到 UIImageView 中的示例函数,该函数应根据需要进行调整。我们建议您在绘制线段时使用圆环,这样就可以将零长度线段绘制为点(例如,将字母 {1} 视为点)。每次写入描边后都会调用 doRecognition() 函数,该函数定义如下。

获取 DigitalInkRecognizer 的实例

为了执行识别,我们需要将 Ink 对象传递给 DigitalInkRecognizer 实例。为了获取 DigitalInkRecognizer 实例,我们首先需要下载所需语言的识别器模型,然后将该模型加载到 RAM 中。为实现此目的,您可以使用以下代码段将其放置在 viewDidLoad() 方法中和使用硬编码语言名称。如需查看示例,了解如何向用户显示可用语言列表并下载所选语言,请参阅快速入门应用

Swift

override func viewDidLoad() {
  super.viewDidLoad()
  let languageTag = "en-US"
  let identifier = DigitalInkRecognitionModelIdentifier(forLanguageTag: languageTag)
  if identifier == nil {
    // no model was found or the language tag couldn't be parsed, handle error.
  }
  let model = DigitalInkRecognitionModel.init(modelIdentifier: identifier!)
  let modelManager = ModelManager.modelManager()
  let conditions = ModelDownloadConditions.init(allowsCellularAccess: true,
                                         allowsBackgroundDownloading: true)
  modelManager.download(model, conditions: conditions)
  // Get a recognizer for the language
  let options: DigitalInkRecognizerOptions = DigitalInkRecognizerOptions.init(model: model)
  recognizer = DigitalInkRecognizer.digitalInkRecognizer(options: options)
}

Objective-C

- (void)viewDidLoad {
  [super viewDidLoad];
  NSString *languagetag = @"en-US";
  MLKDigitalInkRecognitionModelIdentifier *identifier =
      [MLKDigitalInkRecognitionModelIdentifier modelIdentifierForLanguageTag:languagetag];
  if (identifier == nil) {
    // no model was found or the language tag couldn't be parsed, handle error.
  }
  MLKDigitalInkRecognitionModel *model = [[MLKDigitalInkRecognitionModel alloc]
                                          initWithModelIdentifier:identifier];
  MLKModelManager *modelManager = [MLKModelManager modelManager];
  [modelManager downloadModel:model conditions:[[MLKModelDownloadConditions alloc]
                                                initWithAllowsCellularAccess:YES
                                                allowsBackgroundDownloading:YES]];
  MLKDigitalInkRecognizerOptions *options =
      [[MLKDigitalInkRecognizerOptions alloc] initWithModel:model];
  self.recognizer = [MLKDigitalInkRecognizer digitalInkRecognizerWithOptions:options];
}

快速入门应用包含额外的代码,展示了如何同时处理多项下载,以及如何通过处理完成通知来确定下载是否成功。

识别 Ink 对象

接下来是 doRecognition() 函数,为简单起见,从 touchesEnded() 调用该函数。在其他应用中,可能希望在超时后或用户按下按钮触发识别时调用识别。

Swift

func doRecognition() {
  let ink = Ink.init(strokes: strokes)
  recognizer.recognize(
    ink: ink,
    completion: {
      [unowned self]
      (result: DigitalInkRecognitionResult?, error: Error?) in
      var alertTitle = ""
      var alertText = ""
      if let result = result, let candidate = result.candidates.first {
        alertTitle = "I recognized this:"
        alertText = candidate.text
      } else {
        alertTitle = "I hit an error:"
        alertText = error!.localizedDescription
      }
      let alert = UIAlertController(title: alertTitle,
                                  message: alertText,
                           preferredStyle: UIAlertController.Style.alert)
      alert.addAction(UIAlertAction(title: "OK",
                                    style: UIAlertAction.Style.default,
                                  handler: nil))
      self.present(alert, animated: true, completion: nil)
    }
  )
}

Objective-C

- (void)doRecognition {
  MLKInk *ink = [[MLKInk alloc] initWithStrokes:self.strokes];
  __weak typeof(self) weakSelf = self;
  [self.recognizer
      recognizeInk:ink
        completion:^(MLKDigitalInkRecognitionResult *_Nullable result,
                     NSError *_Nullable error) {
    typeof(weakSelf) strongSelf = weakSelf;
    if (strongSelf == nil) {
      return;
    }
    NSString *alertTitle = nil;
    NSString *alertText = nil;
    if (result.candidates.count > 0) {
      alertTitle = @"I recognized this:";
      alertText = result.candidates[0].text;
    } else {
      alertTitle = @"I hit an error:";
      alertText = [error localizedDescription];
    }
    UIAlertController *alert =
        [UIAlertController alertControllerWithTitle:alertTitle
                                            message:alertText
                                     preferredStyle:UIAlertControllerStyleAlert];
    [alert addAction:[UIAlertAction actionWithTitle:@"OK"
                                              style:UIAlertActionStyleDefault
                                            handler:nil]];
    [strongSelf presentViewController:alert animated:YES completion:nil];
  }];
}

管理模型下载

我们已经介绍了如何下载识别模型。以下代码段说明了如何检查模型是否已下载,以及在不再需要模型时如何删除存储空间。

检查模型是否已下载

Swift

let model : DigitalInkRecognitionModel = ...
let modelManager = ModelManager.modelManager()
modelManager.isModelDownloaded(model)

Objective-C

MLKDigitalInkRecognitionModel *model = ...;
MLKModelManager *modelManager = [MLKModelManager modelManager];
[modelManager isModelDownloaded:model];

删除已下载的模型

Swift

let model : DigitalInkRecognitionModel = ...
let modelManager = ModelManager.modelManager()

if modelManager.isModelDownloaded(model) {
  modelManager.deleteDownloadedModel(
    model!,
    completion: {
      error in
      if error != nil {
        // Handle error
        return
      }
      NSLog(@"Model deleted.");
    })
}

Objective-C

MLKDigitalInkRecognitionModel *model = ...;
MLKModelManager *modelManager = [MLKModelManager modelManager];

if ([self.modelManager isModelDownloaded:model]) {
  [self.modelManager deleteDownloadedModel:model
                                completion:^(NSError *_Nullable error) {
                                  if (error) {
                                    // Handle error.
                                    return;
                                  }
                                  NSLog(@"Model deleted.");
                                }];
}

提高文字识别准确性的提示

文字识别的准确性可能因语言而异。准确性还取决于写作风格。虽然数字墨水识别训练经过训练,可以处理多种书写样式,但结果可能因用户而异。

下面是提高文本识别器准确性的一些方法。请注意,这些技巧不适用于表情符号、自动绘制和形状的绘图分类器。

书写区域

许多应用都有明确定义的用户输入区域。符号的含义在一定程度上取决于其大小(相对于包含该符号的写入区域的大小)。例如,大写字母或字母“o”或“c”之间的区别,以及英文逗号和正斜杠。

告知识别器手写区域的宽度和高度有助于提高准确性。不过,识别器会假定写入区域仅包含一行文本。如果物理写入区域足够大,允许用户写入两行或更多行,您可以传入 WriterArea 区域,此时该高度是您对单行文本的最佳估计值,从而可能会获得更好的结果。您传递给识别器的 WRITEArea 对象不必与屏幕上的物理写入区域完全一致。在某些语言中,以这种方式更改 WRITEArea 高度更适合使用某些语言。

指定写入区域时,请指定其宽度和高度(与描边坐标相同)。x,y 坐标参数没有单位要求 - API 会对所有单位进行归一化,因此唯一重要的是描边的相对大小和位置。您可以随意传入适合自己系统的任何缩放比例。

背景信息

上下文是您尝试识别的 Ink 中描边之前紧接着的文本。您可以通过识别前置上下文帮助识别器。

例如,手写字母“n”和“u”经常会被误解。如果用户已输入部分字词“arg”,则系统可能会继续使用可被识别为“ument”或“nment”的描边。指定前置上下文“arg”可解决模棱两可的问题,因为“argument”一词更有可能比“argnment”得到解决。

前置上下文还有助于识别器识别字词间的间隔,即字词之间的空格。您可以输入空格字符,但不能绘制空格,所以识别器如何确定一个字词的结束时间和下一个字词的开始时间?如果用户已经编写了“hello”并继续写了“world”,而没有前置上下文,识别器会返回字符串“world”。但是,如果您指定前置上下文“hello”,则模型将返回字符串“world”,并带有前导空格,因为“hello world”比“helloword”更有意义。

您应提供最长的预上下文字符串,最多 20 个字符(包括空格)。如果字符串较长,则识别器只能使用最后 20 个字符。

以下代码示例展示了如何定义写入区域并使用 RecognitionContext 对象指定前置上下文。

Swift

let ink: Ink = ...;
let recognizer: DigitalInkRecognizer =  ...;
let preContext: String = ...;
let writingArea = WritingArea.init(width: ..., height: ...);

let context: DigitalInkRecognitionContext.init(
    preContext: preContext,
    writingArea: writingArea);

recognizer.recognizeHandwriting(
  from: ink,
  context: context,
  completion: {
    (result: DigitalInkRecognitionResult?, error: Error?) in
    if let result = result, let candidate = result.candidates.first {
      NSLog("Recognized \(candidate.text)")
    } else {
      NSLog("Recognition error \(error)")
    }
  })

Objective-C

MLKInk *ink = ...;
MLKDigitalInkRecognizer *recognizer = ...;
NSString *preContext = ...;
MLKWritingArea *writingArea = [MLKWritingArea initWithWidth:...
                                              height:...];

MLKDigitalInkRecognitionContext *context = [MLKDigitalInkRecognitionContext
       initWithPreContext:preContext
       writingArea:writingArea];

[recognizer recognizeHandwritingFromInk:ink
            context:context
            completion:^(MLKDigitalInkRecognitionResult
                         *_Nullable result, NSError *_Nullable error) {
                               NSLog(@"Recognition result %@",
                                     result.candidates[0].text);
                         }];

描边排序

识别准确性与笔画的顺序密切相关。识别器预期中风的笔画按照自然顺序书写;例如,从左到右表示笔画。任何偏离此模式的情况(例如从最后一个单词开始编写英文句子)都会降低准确度。

另一个示例是移除 Ink 中间的一个单词并将其替换为另一个单词。修订版本可能位于句子的中间,但修订版本的描边位于笔画序列的末尾。 在这种情况下,我们建议您将新写入的字词单独发送到 API,并使用您自己的逻辑将结果与之前的识别值合并。

处理不明确形状

在某些情况下,提供给识别器的形状的含义不明确。例如,边缘非常圆角的矩形可以看作矩形或椭圆形。

这些不明的情况可通过识别识别值(如果有)进行处理。只有形状分类器提供得分。如果模型非常有信心,则排在最前面的结果的得分会远高于次优结果。如果不确定,前两个结果的分数将会接近。另请注意,形状分类器会将整个 Ink 解读为单个形状。例如,如果 Ink 包含一个矩形和一个相邻的椭圆形,则识别器可能会返回一个(或完全不同的一个),因为单个识别候选无法代表两个形状。