Skip to content

Understanding Face Detection and Algorithm

Jump to bottom
K.M.H.Mubin edited this page Jul 4, 2019 · 1 revision

Before we start to write the code first we need to understand how the face detection and algorithm works

How Do Computers “See” Images?

The smallest element of an image is called a pixel, or a picture element. It is basically a dot in the picture. An image contains multiple pixels arranged in rows and columns.

You will often see the number of rows and columns expressed as the image resolution. For example, an Ultra HD TV has the resolution of 3840x2160, meaning it is 3840 pixels wide and 2160 pixels high.

But a computer does not understand pixels as dots of colour. It only understands numbers. To convert colours to numbers, the computer uses various colour models.

In colour images, pixels are often represented in the RGB colour model. RGB stands for Red Green Blue. Each pixel is a mix of those three colours. RGB is great at modelling all the colours humans perceive by combining various amounts of red, green, and blue.

Since a computer only understand numbers, every pixel is represented by three numbers, corresponding to the amounts of red, green, and blue present in that pixel.

In grayscale (black and white) images, each pixel is a single number, representing the amount of light, or intensity, it carries. In many applications, the range of intensities is from 0 (black) to 255 (white). Everything between 0 and 255 is various shades of grey.

If each grayscale pixel is a number, an image is nothing more than a matrix (or table) of numbers:


Fig: Pixel Matrix

In color images, there are three such matrices representing the red, green and blue channels.

What Is Face Detection?

Face detection is a computer technology that is being applied for many different applications that require the identification of human faces in digital images or video.

This is very easy for humans, but computers need precise instructions. The images might contain many objects that aren’t human faces, like buildings, cars, animals, and so on. It is distinct from other computer vision technologies that involve human faces, like facial recognition, analysis, and tracking.

How does Face Detection work?

Firstly the image is imported by providing the location of the image. Then the picture is transformed from RGB to Grayscale because it is easy to detect faces in the grayscale. After that, the image manipulation used, in which the resizing, cropping, blurring and sharpening of the images done if needed. The next step is image segmentation, which is used for contour detection or segments the multiple objects in a single image so that the classifier can quickly detect the objects and faces in the picture. The next step is to use the algorithm. The algorithm used for finding the location of the human faces in a frame or image. All human faces shares some universal properties of the human face like the eyes the region is darker than its neighbour pixels and nose region is brighter than the eye region.

Face Recognition Algorithm

Over the last ten years or so, face recognition has become a popular area of research in computer vision and one of the most successful applications of image analysis and understanding. There are different types of Traditional face recognition algorithms, for example:

• Eigenfaces (1991)

• Local Binary Patterns Histograms (LBPH) (1996)

• Fisher faces (1997)

• Scale Invariant Feature Transform (SIFT) (1999)

• Speed Up Robust Features (SURF) (2006)

Each method has a different approach to extract the image information and perform the matching with the input image. However, the methods Eigenfaces and Fisher faces have a similar approach as well as the SIFT and SURF methods. One of the oldest (not the oldest one) and more popular face recognition algorithms: Local Binary Patterns Histograms (LBPH).

Local Binary Patterns Histograms (LBPH) (1996)

Local Binary Pattern (LBP) is a simple yet very efficient texture operator which labels the pixels of an image by thresholding the neighbourhood of each pixel and considers the result as a binary number.

It was first described in 1994 (LBP) and has since been found to be a powerful feature for texture classification. It has further been determined that when LBP is combined with histograms of oriented gradients (HOG) descriptor, it improves the detection performance considerably on some datasets. Using the LBP combined with histograms we can represent the face images with a simple data vector.

How LBPH algorithm work step by step:

LBPH algorithm work in 5 steps.

  1. Parameters: the LBPH uses 4 parameters:

  • Radius: the radius is used to build the circular local binary pattern and represents the radius around the central pixel. It is usually set to 1.

  • Neighbors: the number of sample points to build the circular local binary pattern. Keep in mind: the more sample points you include, the higher the computational cost. It is usually set to 8.

  • Grid X: the number of cells in the horizontal direction. The more cells, the finer the grid, the higher the dimensionality of the resulting feature vector. It is usually set to 8.

  • Grid Y: the number of cells in the vertical direction. The more cells, the finer the grid, the higher the dimensionality of the resulting feature vector. It is usually set to 8.

  1. Training the Algorithm: First, we need to train the algorithm. To do so, we need to use a dataset with the facial images of the people we want to recognize. We need to also set an ID (it may be a number or the name of the person) for each image, so the algorithm will use this information to recognize an input image and give you an output. Images of the same person must have the same ID. With the training set already constructed, let’s see the LBPH computational steps.

  2. Applying the LBP operation: The first computational step of the LBPH is to create an intermediate image that describes the original image in a better way, by highlighting the facial characteristics. To do so, the algorithm uses a concept of a sliding window, based on the parameters radius and neighbours.

The image below shows this procedure:


Fig: LBP Operation

Based on the image above, let’s break it into several small steps so we can understand it easily:

  • Suppose we have a facial image in grayscale.

  • We can get part of this image as a window of 3x3 pixels.

  • It can also be represented as a 3x3 matrix containing the intensity of each pixel (0~255).

  • Then, we need to take the central value of the matrix to be used as the threshold.

  • This value will be used to define the new values from the 8 neighbours.

  • For each neighbour of the central value (threshold), we set a new binary value. We set 1 for values equal or higher than the threshold and 0 for values lower than the threshold.

  • Now, the matrix will contain only binary values (ignoring the central value). We need to concatenate each binary value from each position from the matrix line by line into a new binary value (e.g. 10001101). Note: some authors use other approaches to concatenate the binary values (e.g. clockwise direction), but the final result will be the same.

  • Then, we convert this binary value to a decimal value and set it to the the central value of the matrix, which is actually a pixel from the original image.

  • At the end of this procedure (LBP procedure), we have a new image which represents better the characteristics of the original image.

Fig: LBP Bilinear Interpolation

It can be done by using bilinear interpolation. If some data point is between the pixels, it uses the values from the 4 nearest pixels (2x2) to estimate the value of the new data point.

  1. Extracting the Histograms: Now, using the image generated in the last step, we can use the Grid X and Grid Y parameters to divide the image into multiple grids,as can be seen in the following image:

Fig: Extracting The Histogram

Based on the image above, we can extract the histogram of each region as follows:

  • As we have an image in grayscale, each histogram (from each grid) will contain only 256 positions (0~255) representing the occurrences of each pixel intensity.

  • Then, we need to concatenate each histogram to create a new and bigger histogram. Supposing we have 8x8 grids, we will have 8x8x256=16.384 positions in the final histogram. The final histogram represents the characteristics of the image original image.

  1. Performing the face recognition: In this step, the algorithm is already trained. Each histogram created is used to represent each image from the training dataset. So, given an input image, we perform the steps again for this new image and creates a histogram which represents the image.

  • So to find the image that matches the input image we just need to compare two histograms and return the image with the closest histogram.

  • We can use various approaches to compare the histograms (calculate the distance between two histograms), for example: Euclidean distance, chi-square, absolute value, etc. In this example, we can use the Euclidean distance (which is quite known) based on the following formula:


Fig: Euclidean distance

  • So the algorithm output is the ID from the image with the closest histogram. The algorithm should also return the calculated distance, which can be used as a ‘confidence’ measurement. Note: don’t be fooled about the ‘confidence’ name, as lower confidences are better because it means the distance between the two histograms are closer.

  • We can then use a threshold and the ‘confidence’ to automatically estimate if the algorithm has correctly recognized the image. We can assume that the algorithm has successfully recognized if the confidence is lower than the threshold defined.

### References

Clone this wiki locally