Locating, measuring, or recognising an object seen in an image requires finding what pixels in the image correspond to that object. This set of lectures introduces several techniques for locating structures: exploiting differences in colour or lightness, differences in positions between two consecutive frames, or grouping by colour. The lectures also introduce a widely used family of image processing operators: 2D convolution.
This lecture shows a few examples of region segmentation, where the image has been partitioned into meaningful regions. It also introduces several approaches to segmentation.
The simplest approach to segmentation uses a threshold value to separate the desired region or regions from the rest, based on whether pixel values are larger or smaller than the given threshold value. This lecture introduces this concept and also introduces a method to select the threshold value. The selection process introduces the concept of a histogram and also introduces convolution, which here is used to smooth the histogram.
Convolution is a widely used image processing technique, consisting of summing together a set of values multiplied by a set of weights. Here, the input values are from image patches surrounding a central point. The weights can be chosen for a variety of effects, such as noise reduction or feature enhancement.
More complex scenes make simple thresholding difficult because it is hard to tell the objects of interest from the rest of the image. If the camera and background scene are fixed, then new objects can be detected because they are different from the background, which is the focus of this lecture. This is applied to color images. We also introduce normalized RGB and a ratio technique that compensates for extreme lighting variations.
This lecture introduces an approach to segmenting the interesting regions in the image based on adjacent pixels that have similar colors. Colors are grouped into sets of similar colors using mean-shift clustering, which is applied in La*b* color space rather than in the image space.