Research on the Image Segmentation based on Improved Region Extractions

Yi Zhang, Xiuting YangProceedings of the 2018 International Conference on Network, Communication, Computer Engineering (NCCE 2018) 2018 · DOI

Abstract

Image segmentation is one of the classic methods of image signal processing, widely used in scientific research, biomedical engineering, and military guidance. Based on region extraction, this paper introduces and compares three typical image segmentation methods: the region growth algorithm, the watershed algorithm, and the split-and-merge method. Computer simulations and comparative results show that while the region growth method can handle complex image segmentation, its time and space costs are high, leading to low efficiency. The watershed method effectively separates adhered objects in an image but is significantly influenced by noise, often resulting in over-segmentation. The effects of the split-and-merge method and the region growth method are found to be similar, with no significant difference between them.

Introduction

Image segmentation serves as a fundamental challenge in computer vision, acting as the critical bridge between raw pixel data and high-level object recognition. By partitioning a digital image into multiple segments, researchers can simplify representations for easier analysis. This study focuses specifically on region-based extraction techniques. Unlike edge-detection methods that seek discontinuities, region-based methods identify groups of pixels sharing commonalities such as texture or gray-level intensity. The core objective is to navigate the trade-off between segmentation accuracy—particularly for "adhered" or touching objects—and computational efficiency.

Methodology

The study evaluates three primary algorithmic frameworks for region-based segmentation:

  1. Region Growth Method: A "bottom-up" approach that begins with a seed pixel and examines neighbors using 4- or 8-connectivity. Pixels meeting a similarity criterion (such as gray-level difference below a threshold) are merged into the region until no further pixels can be added.
  2. Watershed Method: This method treats the image as a topographic map where pixel intensity represents elevation. It utilizes two main implementations: the Rainfall Method, which simulates water flow to find catchment basins, and the Submergence Method, which simulates flooding from local minima. Watershed lines (dams) are constructed where different basins meet to separate regions.
  3. Split-Merge Method: A hybrid "top-down" and "bottom-up" approach. It recursively splits the image into quadrants if they are inconsistent (high variance) and subsequently merges adjacent quadrants that meet similarity criteria.

Experimental Setup

The algorithms were tested using two specific simulation scenarios to evaluate performance under different conditions:

  • Scenario 1: A high-contrast image of an aircraft against a bright, slightly noisy sky background.
  • Scenario 2: A complex biological image showing blood cells that are "adhered" or touching one another.
Original test image of an aircraft against a sky background used to evaluate segmentation performance.

Results and Discussion

Quantitative results highlight a significant efficiency gap between the methods. The Watershed method was the fastest, processing in approximately 0.435 seconds. In comparison, Region Growth was the slowest at 1.660 seconds—roughly 2 to 4 times slower than the Watershed approach. For context, Local Thresholding and Otsu's Method recorded times of 0.833s and 1.346s, respectively.

Qualitatively, the Watershed method proved superior for separating adhered objects. By identifying the topographic "ridge" between two basins, it successfully delineated individual cells that other methods failed to separate clearly.

Segmentation of adhered blood cells demonstrating the Watershed method's ability to separate touching objects.

However, the Watershed method's primary weakness is its extreme sensitivity to image noise. Even minor fluctuations can create "fake" local minima, leading to over-segmentation, where a single object is fractured into many meaningless pieces. Conversely, the Split-Merge and Region Growth methods produced functionally equivalent results, suggesting that the choice between them often depends on implementation preference rather than output quality.

Conclusion

This comparative study underscores the distinct trade-offs in region-based segmentation. Region Growth is conceptually robust but computationally heavy, limiting its use in real-time environments. The Watershed method offers high speed and unique capabilities for separating touching objects but requires significant noise-reduction pre-processing. Future work in "Improved Region Extractions" should aim to develop hybrid methods that leverage the efficiency of Watershed while incorporating the noise-resilience and similarity criteria of growth-based techniques.

Summary updated

Abstract

Image segmentation is one of the classic methods of image signal processing. It is widely used in the scientific research, biomedical engineering, military guidance and so on. Based on the region extraction, this paper introduces three typical image segmentation methods: region growth algorithm, watershed algorithm and split and merge method. Computer simulation and the comparison of results show that, Region growth method can be used to break up complex image segmentation, but the cost of time and space are very high, the efficiency is low. By using watershed segmentation, adhesion of the image can be separated, but a watershed method greatly influenced by the noise, easy to excessive segmentation. The effects of split and merge segmentation method and region growth method are similar; there is no significant difference between them.
BibTeX
@inproceedings{zhang2018research,
  author = {Yi Zhang and Xiuting Yang},
  title = {{Research on the Image Segmentation based on Improved Region Extractions}},
  booktitle = {Proceedings of the 2018 International Conference on Network, Communication, Computer Engineering (NCCE 2018)},
  year = {2018},
  doi = {10.2991/ncce-18.2018.74},
}
Figure