Adaptive Local Thresholds for Still Color Image Segmentation
This research presents an adaptive segmentation algorithm for still color images. The key idea is that a single global threshold is insufficient for complex images; instead, the algorithm derives multiple local thresholds automatically during an iterative merging process. The thresholds are not predefined: their number K and values t₁,...,tₖ are discovered from the image itself.
First Stage — Watershed Over-Segmentation
The image I is converted to a gradient image Gᵢ, and the watershed transform generates an initial partition {Rᵢ⁰}. This intentionally produces many primitive regions, giving the algorithm a fine initial structure from which reliable larger regions can be formed.
Second Stage — RAG and MST-Driven Merging
The partition is represented by a Region Adjacency Graph G=(V,E), where each node is a region and each edge connects adjacent regions. Edge weights are defined by a dissimilarity function f(Rᵢ,Rⱼ) combining hue-color difference and boundary-gradient information. The merging order follows a Kruskal-like MST process, merging the most similar adjacent regions first.
Third Stage — Iterative Local Threshold Discovery
During merging, the algorithm monitors the homogeneity of each growing region through the variance of the V component in HSV space, denoted σ(Rᵢᵐ). A local threshold is derived when the homogeneity change Δσ(Rᵢᵐ) reaches the first meaningful local maximum that indicates an inhomogeneous merge.
Fourth Stage — Regression and Final Regions
Once a threshold tₛ is detected, the algorithm performs a regression process: it reverses merges from the end of the merging tree until the merge that created the inhomogeneous region is removed. The two separated regions are marked as final regions. The process repeats until no mergeable regions remain, yielding an adaptive segmentation tailored to the specific image.