The limitations of thresholding are usually related to data inconsistencies contained in the image. Sometimes from a global perspective, an image might seem like it is easily differentiated into object and background, like the image shown in the example above. But look at the original image a little closer, and problems arise.
One of the core problems is related to non-uniform regions. This usually manifests itself in contiguous regions which are not one continuous tone of gray, i.e. non-uniform. Consider the image below (extracted from the plane spotters card) In the contiguous regions such as the plane and background, there are discontinuities. In this case those defects aren’t that problematic, but often there are images in which the differential between object and background is not that high.
Here is an example of how non-uniform a contiguous region can be:
Lack of Step-edges
When segmenting an object from the background regions, clean, crisp step-edges are a dream to work with. What are step edges? Exactly what they appear to be, edges that are almost vertical in nature from one gray tone to another. The edges in the plane seem fairly crisp, until you look closer and realize that they are not – they appear more jagged. Curves on objects that seemed crisp as well are not, and that is a trick of the eyes, which at a lower resolution, make the curves seem smooth. Here is an example:
It is not until the regions are enlarged that the problems manifest themselves. The jagged contours with literally turn into jagged edges, once the binarization process is complete.
Images that are very complex will be hard to binarize effectively. If there is more than one object in the image, then using a 2-level thresholding algorithm may not be that useful. It may be possible to use a more complex segmentation technique, or possibly explore the use of colour.
How to fix these issues?
With some images it is possible to fix intensity discontinuities through the application of some form of edge-preserving filter, for example Kuwahara filter. If there is extremely small detail it may also be necessary to improve the resolution of the image.