Image segmentation based on Blob analysis and quad-tree algorithm
Wen‐quan Fan, Wen‐shu Xiao · 2018
Image segmentation is one of the most popular topic in recent research and studies, there are lots of different method to solve this problem. Some existing method works pretty well, and others not. This paper proposed to implement two improved existing method, Split and merge and Blob coloring algorithm, and compared their segmentation results in both 2D and 3D. Meanwhile, to clarify our achievements following scientific method, we have to establish our evaluation function FMI to give a score to tell what level of goodness our implementation could achieve. The first task is implementation of different region growing algorithms. We implement a general split and merge algorithm and the Blob Coloring algorithm that can work both in 2D and in 3D with different homogeneity criteria. For the split and merge algorithm, we will use quad-tree algorithm to split and merge our image based on the homogeneity. For the blob algorithm, we will check the L shape, then merge the pixels with similar homogeneity, and ignore those not. According to the experimental results, We found the split merge algorithm is generally better than the blob algorithm. Although for some case, the blob algorithm can also reach to more than 0.8 FMI score in evaluating, the overall performance is still bad. Split and merge algorithm works well on all the image cases. As we think, we decide to apply our implementation in some real images to achieve some goals. For example, we choose a group of interior design and a group of animals as 3D images. The result coming from our implementation could be used in biomedical field such as cancer detection, as well as in public management such as suspects outline detection. However, there are real problems for us to use region growing techniques to detect boundary of objects with some specific indexes because if the target has obvious difference inside, the algorithm will not treat it as a whole. However, from a visual point of view, we treat it as a complete object.