Image processing (2) - image feature extraction LBP
In fact, everyone is talking about image processing now. In fact, computers do not understand pictures. The reason why they can process images is that images are actually matrices, which are actually numbers. In fact, they are all processing numbers. It is not surprising that deep learning has made great achievements in images. The image feature extraction methods that you have heard before include HOG, Haar, LBP, SIFT, etc., but now they are gradually fading out of everyone's field of vision. They are replaced by deep learning. Visual deep learning exists A very difficult problem-similar to a black box, everyone does not know the process, only knows the input and output, which is very unfriendly to researchers. Today at noon, my brother released a new open source tool that can visualize the middle part of the network , I will continue to update it later. Today I mainly talk about LBP
1. Background
LBP (Local Binary Pattern): is an operator that describes the local features of an image
2. Basic version of LBP
- Each pixel in each image can be described by surrounding pixels. Each pixel can have a corresponding LBP value
- Now take a 3*3 matrix, and the LBP value of the pixel in the middle of the matrix is calculated from the eight pixels around it
- If it is greater than the middle pixel point, the value is 1, otherwise it is 0, and then it is read in clockwise order to be an 8-bit binary number, and after converting it into decimal, it is the LBP value of the point, and other points are calculated similarly
- Example: The LBP value of the center point of the following matrix is finally 37
The above is the simplest and most basic LBP process. Later, people improved it, and the following appeared:
3. Round LBP
In order to adapt to texture features of different scales and meet the requirements of grayscale and rotation invariance, Ojala et al. improved the LBP operator, extended the 3×3 neighborhood to any neighborhood, and replaced the square neighborhood with a circular neighborhood. area
- You can specify the radius of the domain and the number of adjacent pixels
- Robust to illumination changes
- The smaller the number of pixels in the field, the lower the brightness of the image
- The smaller the field radius, the clearer the texture of the image (equivalent to pixel-by-pixel processing)
- example:
4. Rotation invariance
The circular LBP has gray scale invariance, but does not have rotation invariance. After the starting position is determined, the solution: by rotating the obtained LBP features, a series of LBP feature values are obtained, and finally the smallest feature mode is used as LBP feature of the center pixel
There are also MB-LBP features - average gray value, SEMB-LBP, etc., to understand by yourself