iOS-Vision Framework - Intro

Intro To Vision :

Vision was introduced in 2017 WWDC along with list of other machine learning frameworks apple released (Core ML,NLP).Vision can be used on both image as well as sequences of image (videos).We can also integrate vision with Core ML Models for example it can used to give Core ML required input parameters example for MINST image classification we can detect numbers as rect in image using vision and send it Core ML model for prediction. Check the WWDC Video

Vision Main Features

• Face Detection and Recognition
• Machine Learning Image Analysis
• Barcode Detection
• Image Alignment Analysis
• Text Detection
• Horizon Detection
• Object Detection and Tracking

Project Overview

In this project we are gonna look into simple image analysis techniques like marking Rectangle objects,faces ,text boxes,Char on texts

For performing any image analysis operation we need follow this three step process

• Create a Vision Image Request

• Create a Image Request Handler

• Assigning Image Requests To Request handler

Face Detection

• Creating face detection request

   lazy var faceDetectionRequest : VNDetectFaceRectanglesRequest = {
        let faceRequest = VNDetectFaceRectanglesRequest(completionHandler:self.handleFaceDetection)
        return faceRequest
    }()

• Create a image request handler

        let handler = VNImageRequestHandler(ciImage: ciImage, orientation: Int32(uiImage.imageOrientation.rawValue))

• Assigning image requests to request handler

     try handler.perform([self.faceDetectionRequest])

• Handler code

         func handleFaceDetection (request: VNRequest, error: Error?) {
        guard let observations = request.results as? [VNFaceObservation]
            else { print("unexpected result type from VNFaceObservation")
                return }
        guard observations.first != nil else {
            return
        }
        // Show the pre-processed image
        DispatchQueue.main.async {
            self.analyzedImageView.subviews.forEach({ (s) in
                s.removeFromSuperview()
            })
            for face in observations
            {
                let view = self.CreateBoxView(withColor: UIColor.red)
                view.frame = self.transformRect(fromRect: face.boundingBox, toViewRect: self.analyzedImageView)
                self.analyzedImageView.image = self.originalImageView.image
                self.analyzedImageView.addSubview(view)
                self.loadingLbl.isHidden = true
                
            }
        }
    }

• Converting vision rect to uikit rect

  one main thing to keep in mind that vision rect values are different from others

  Vision: origin ---> bottom left

  Size ---> Max value of 1

  UIkit : origin ---> top left

  Size ---> UIVIEW bounds

  
     //Convert Vision Frame to UIKit Frame
    func transformRect(fromRect: CGRect , toViewRect :UIView) -> CGRect {
  
        var toRect = CGRect()
        toRect.size.width = fromRect.size.width * toViewRect.frame.size.width
        toRect.size.height = fromRect.size.height * toViewRect.frame.size.height
        toRect.origin.y =  (toViewRect.frame.height) - (toViewRect.frame.height * fromRect.origin.y )
        toRect.origin.y  = toRect.origin.y -  toRect.size.height
        toRect.origin.x =  fromRect.origin.x * toViewRect.frame.size.width
  
        return toRect
    }
  
  

• Box view

For drawing rectangle box around our detection

      func CreateBoxView(withColor : UIColor) -> UIView {
        let view = UIView()
        view.layer.borderColor = withColor.cgColor
        view.layer.borderWidth = 2
        view.backgroundColor = UIColor.clear
        return view
    }

CHAR DETECTION

• Creating char detection request

        lazy var textRectangleRequest: VNDetectTextRectanglesRequest = {
        let textRequest = VNDetectTextRectanglesRequest(completionHandler: self.handleTextIdentifiaction)
        textRequest.reportCharacterBoxes = true
        return textRequest
    }()
  

• Create a image request handler

        let handler = VNImageRequestHandler(ciImage: ciImage, orientation: Int32(uiImage.imageOrientation.rawValue))
  

• Assigning image requests to request handler

       try handler.perform([self.textRectangleRequest])
  

• Handler code

      func handleTextIdentifiaction (request: VNRequest, error: Error?) {
  
      guard let observations = request.results as? [VNTextObservation]
          else { print("unexpected result type from VNTextObservation")
              return
          }
      guard observations.first != nil else {
          return
      }
      DispatchQueue.main.async {
          self.analyzedImageView.subviews.forEach({ (s) in
              s.removeFromSuperview()
          })
          for box in observations {
              guard let chars = box.characterBoxes else {
                  print("no char values found")
                  return
              }
              for char in chars
              {
                  let view = self.CreateBoxView(withColor: UIColor.green)
                  view.frame = self.transformRect(fromRect: char.boundingBox, toViewRect: self.analyzedImageView)
                  self.analyzedImageView.image = self.originalImageView.image
                  self.analyzedImageView.addSubview(view)
                  self.loadingLbl.isHidden = true
              }
          }
      }
  
  }  
  

RECTANGLE DETECTION

• Creating rectangle detection request

       lazy var rectangleBoxRequest: VNDetectRectanglesRequest = {
      return VNDetectRectanglesRequest(completionHandler:self.handleRectangles)
  }()

• Create a image request handler

        let handler = VNImageRequestHandler(ciImage: ciImage, orientation: Int32(uiImage.imageOrientation.rawValue))
  

• Assigning image requests to request handler

        try handler.perform([self.rectangleBoxRequest])
  

• Handler code

       func handleRectangles(request: VNRequest, error: Error?) {
  
      guard let observations = request.results as? [VNRectangleObservation]
          else { print("unexpected result type from VNDetectRectanglesRequest")
                  return
      }
      guard observations.first != nil else {
          return
      }
      // Show the pre-processed image
      DispatchQueue.main.async {
          self.analyzedImageView.subviews.forEach({ (s) in
              s.removeFromSuperview()
          })
          for rect in observations
          {
              let view = self.CreateBoxView(withColor: UIColor.cyan)
              view.frame = self.transformRect(fromRect: rect.boundingBox, toViewRect: self.analyzedImageView)
              self.analyzedImageView.image = self.originalImageView.image
              self.analyzedImageView.addSubview(view)
              self.loadingLbl.isHidden = true
          }
      }
  }
  

You can also group Requests and Perform analysis

    handler.perform([self.textRectangleRequest,self.faceDetectionRequest,self.rectangleBoxRequest])

Here am performing text,face and rectangle box analysis on a single input