Deep Learning Face Recognition: Part 2

-
Before reading this, please read deep-learning-face-recognition-part-1.html for better understanding

 What is Face Detection? (locate and extract faces from each image.)
The ability to detect and locate human faces in a photograph.
Face detection is used for various purposes but here we use it for the only extraction of faces and pass it to the next level of face recognition. (face detection is one of the steps in our pipeline)
Easiest Way to detect face is applying Sliding Window Classifier, It is done in two steps:
  1. Build a face detection model using a machine learning model which can tell whether a given image is a face or not.
  2. Slide the face detector across the large image for faces. If faces are detected it will note the location of it.
Popular face detection algorithms:(Accuracy of detection increase from 1 to 3)
  1. viola jones
    1. Invented by Paul Viola and Michael Jones in the early 2000s.
    2. Uses decision trees to detect faces based on light and dark areas.
    3. Very fast and great for low powered devices.
    4. Not very accurate
  2. Histograms of oriented gradients(HOG)
    1. Invented in 2005
    2. Looks for the shift from light to dark areas in an image.
    3. Slower than Viola-Jones, but more accurate.
    4. runs well on normal computers without special software.
  3. Convolutional neural network
    1. Uses a deep neural network to detect faces.
    2. Very accurate, but requires a lot of training data.
    3. Runs best on computers with dedicated GPUs
Here I am using the HOG algorithm, so I will explain HOG.
  1. Convert the given image into black and white as the HOG algorithm looks for only white and dark areas.
  2. Now we start looking at every single pixel of the image and start measuring which surrounding pixel of it is much darker than it. Then we allocate that direction to it and mark an arrow sign in that direction on that pixel.
  3. After completing every pixel of the image we get a large number of lines on the image which indicates gradients. Each gradient shows how the light flows from a dark area to dark area
    original face image

    The gradient of each pixel in the image
  4. It's a huge number of gradients in the above image. So to reduce them take a small area in the above image and replace that area by the maximum frequency of directions in that area
    the small area selected image of above
    replaced by maximum count direction of the directions in that area
  5. Repeat the above process until the complete image is processed.
    final Image of HOG
Finding faces in images with HOG features:
  1. Collecting training data and converting them to HOG representations of faces.
  2. By using HOG representations of faces we train a machine learning model. (HOG can perform well even with small dataset)
  3. Sliding window classifier on HOG representation of the new image and try to find the face.
why Use Hog?
  1. HOG is a simplified representation of an image that still captures enough details to detect faces
  2. HOG representation is not affected by small changes in lighting.
  3. A HOG representation is not affected by small changes in an object's shape
Now moving to code section of 1:

Output:
If we zoom in for getting each pixel view, we get it as follows:


Training and sliding window classifier is discussed in the next part.


Comments

Post a Comment

Popular posts from this blog

Splitting criteria

-
Image
This package helps to find splitting criteria in C4.5 Information Gain ID3 uses information gain as its attribute selection measure. This measure is based on pioneering work by Claude Shannon on information theory, which studied the value or “information content” of messages. Let node N represent or hold the tuples of partition D. The attribute with the highest information gain is chosen as the splitting attribute for node N. This attribute minimizes the information needed to classify the tuples in the resulting partitions and reflects the least randomness or “impurity” in these partitions. Such an approach minimizes the expected number of tests needed to classify a given tuple and guarantees that a simple (but not necessarily the simplest) tree is found. The expected information needed to classify a tuple in D is given by: where pi is the probability that an arbitrary tuple inDbelongs to classCi and is estimated by |Ci,D|/|D|. A log function to the base 2 is used becaus
Read more

QUOTATION SERVER AND CLIENT

-
Image
A simple program of server and client where the server sent a good quote to the connected client immediately after the connection(as welcoming). Multiple Clients are handled using java multithreading concept, Used UDP datagram sockets, and text where quotes are stored.  The output of the above program: Outputs after quotes are completed:
Read more

How to Handle Noisy Data in preprocessing of data?

-
Binning method:(one of the method) first sort data and partition into (equi-depth) bins then one can smooth by bin means,  smooth by bin median, smooth by bin boundaries, etc. Equal-width (distance) partitioning: It divides the range into N intervals of equal size: uniform grid if A and B are the lowest and highest values of the attribute, the width of intervals will be: W = (B-A)/N. The most straightforward But outliers may dominate the presentation Skewed data is not handled well. Equal-depth (frequency) partitioning: It divides the range into N intervals, each containing approximately same number of samples Good data scaling Managing categorical attributes can be tricky. Code for binning (if needed we can edit for user input instead of random) : Feel free to comment about mistakes and doubts.
Read more