Introduction to Probability

Multivariable Calculus

Algorithms: Part II

Algorithms: Part I

Introduction to Software Design and Architecture

Calculus Two: Sequences and Series

LAFF Linear Algebra

Stanford Machine Learning

Calculus One

Computational Thinking

Effective Thinking Through Mathematics

CS50 Introduction to Computer Science


Week 11 - Application Example: Photo OCR

Problem Description and Pipeline

  • Photo OCR: Photo Optical Character Recognition.
  • Photo OCR Pipeline(a machine learning pipeline):
  • Pipelines are common terms in machine learning
    • Separate modules which may each be a machine learning component or data processing component

Sliding Windows Classifier

  • In order to talk about detecting things in images let’s start with a simpler example of pedestrian detection.

Supervised Learning for Pedestrian Detection

  • Given the training set:

    • xx = pixels in 82x36 image patches
  • Now we have a new image - how do we find pedestrians in it?

    • * Start by taking a rectangular 82x36 patch in the image * Keep stepping rectangle along all the way to the right with 4 pixels/step(always 5-8 pixels). * Then move back to the left hand side but step down a bit too. * Keep the steps until the last line.
    • * Now we initially start with a larger image patch (of the same aspect ratio) * Each time we process the image patch, we're resizing the larger patch to a smaller image, then running that smaller image through the classifier.
    • Hopefully, we will eventually get this:
  • Back to Text Detection

    • Like pedestrian detection, we generate a labeled training set with
      • Positive examples (some kind of text)
      • Negative examples (not text)
    • Having trained the classifier we apply it to an image
      • So, run a sliding window classifier at a fixed rectangle size
      • If you do that end up with something like this
        • Black - no text
        • White - text
      • For text detection, we want to draw rectangles around all the regions where there is text in the image
      • Take classifier output and apply an expansion algorithm
        • Takes each of white regions and expands it
      • Look at connected white regions in the image above

Character Segmentation

  • Look in a defined image patch and decide, is there a split between two characters?
  • Train a classifier to classify between positive and negative examples
  • Use a 1-dimensional sliding window to move along text regions
    • Does each window snapshot look like the split between two characters?
      • If yes insert a split
      • If not move on

Character Classification

  • Multi-class characterization problem

Getting Lots of Data: Artificial Data Synthesis

Artificial Data Synthesis for Photo OCR

  1. Use computer’s font library, or online font libraries. Take different fonts, paste them with random backgrounds
  2. Distort the exist data set
  • Synthesizing data by introducing distortions: Speech recognition
    • We can add noisy background to the original audio to make it unclear

Getting More Data

  • When do we need to get more data?
    • Make sure we have a low bias classifier. (Plot learning curves)
  • When we really need it, ask ourselves: “How much work would it be to get 10x as much data as we currently have?”
    • Artificial data synthesis
      ­* Collect/label it yourself
    • “Crowd source” (E.g. Amazon Mechanical Turk)

Ceiling Analysis

  • Estimating the errors due to each component.
  • Decide what part of the pipeline we should spend the most time trying to improve.
  • Take the Photo OCR pipeline as the example:
    • We find that our test set has 72% accuracy.
    • Steps:
      1. Go to the first module - Text detection. Manually tell the algorithm where the text is.
        • Simulate if your text detection system was 100% accurate
        • Check how this change affects the accuracy of the overall system.
        • Accuracy goes up to 89%
      2. Next do the same for the character segmentation
        • Accuracy goes up to 90% now
      3. Finally doe the same for character recognition
        • Goes up to 100%
    • Base on the analysis, we know which module to improve.