Object detection with Haar Cashcode Classifier

To briefly introduce you about the Haar Cashcode,

- The method was proposed by P.Viola and M.Johns in 2001

- Machine learning method

- trained on both positive and negative images

1. positive image: includes the object we are trying to detect

2. negative image: the background

Based on the Haar Wavelet sequence

Haar Wavelet: convolution kernels used to extract features

*Haar like features: Edges, Lines, Diagonal edges

First, the integral image concept

= each pixel represents the cumulative sum of the corresponding input pixels above and to the left of the pixel

ex) The viola-johns paper used a 24 by 24 base window size and that would result in more that 180,000 features calculated in the integral image

Second, An AdaBoost classifier is used to reduce the number of features

AdaBoost: the algorithm selects a few important features from a large set to give highly efficient classifiers by employing the AdaBoost

1. A weak classifier is made on top of the training data based on the weighted samples

2. It selects only those features that help to improve the classifier accuracy

3. AdaBoost cuts down the number of features - originally over 180,000 features were generated but it cuts down to about 6,000 features by using the AdaBoost

ex) Cats and Dogs

each weak classifier splits the examples with at least 50% accuracy

- the misclassified examples are then emphasized on the next round

To set weights to both classifiers and samples in a way that forces classifiers to concentrate on observations that have been misclassified

---> The process is repeated until it has minimized the number of error and contracts a strong classifier

Third, A cascade of classifiers

= This classifier group's sub-images from the input image in stages and disregards any region that doesn't match the object it is trying to detect

*The classifier groups the features into multiple sub-images and the classifier at each stage determines whether the sub-image is the object we are trying to detect

- In the case that it is not, the sub-windows is discarded only with the features in the window

- If the sub-window moves past the classifier, it continues to the next stage where the second stage of features is applied

Copyright Coursera All rights reserved