Image Features

The image intensities for classification does not function well

Therefore, we convert the image to vector

- use large patches for illustrative purposes instead of individual pixel values

Classifying an image involves the relationship between pixels

Features are measurements taken from the image that help with classification (using histogram)

Histogram only counts intensities and does not consider the relationship between pixels

The histogram only quantified the fact that they have the same number of black pixels

--> To overcome this problem, split image into sub-image and calculate the histogram for each sub-image

Color is not always the best tool for classification

- Red color

elements of red channel would have large values

- Blue color

only the elements of the blue or green channel would have a large magnitude

Although human use colors to classify images, it's hard for people to come up with features the use colors

--> Gradients look identical

Therefore, we use features based on gradients

H.O.G

- one of many image features we can use

- Histogram of Oriented Gradient

- H.O.G uses the gradient orientation of the located regions of the image

- H.O.G generates a histogram for each localized region

How H.O.G works?

We can calculate the histogram where each sin index is the angle of the gradient, the value of the histogram is the magnitude

General H.O.G is calculated as follows assuming we have an image of a kitten

- convert the image to grayscale

- calculate the magnitude and angles of the gradients using Sobel

The images are divided in a grid fashion into cells, and for the pixels within each cell, a histogram of gradient directions is compiled

H.O.G feature vector is a combination of all pixel-level histogram and used with sum to classify the image

Summarize the entire machine learning process

Feature extraction (H.O.G) -----> kernel (non-linear mapping) -----> linear classification