[Paper Review] Image Segmentation using Multi-Threshold technique by Histogram Sampling

Amit Gurung, Sangyal Lama Tamang|arXiv (Cornell University)|Sep 11, 2019

Advanced Image and Video Retrieval Techniques33 references4 citations

TL;DR

This paper proposes AMTIS, a heuristic multi-threshold image segmentation method that automatically selects optimal thresholds by sampling the histogram's valleys and partitioning it into equal regions to ensure uniform distribution. The approach achieves up to 35.58× speedup over Otsu’s method while maintaining or improving segmentation quality, as validated by PSNR, SSIM, and FSIM metrics on standard benchmarks.

ABSTRACT

The segmentation of digital images is one of the essential steps in image processing or a computer vision system. It helps in separating the pixels into different regions according to their intensity level. A large number of segmentation techniques have been proposed, and a few of them use complex computational operations. Among all, the most straightforward procedure that can be easily implemented is thresholding. In this paper, we present a unique heuristic approach for image segmentation that automatically determines multilevel thresholds by sampling the histogram of a digital image. Our approach emphasis on selecting a valley as optimal threshold values. We demonstrated that our approach outperforms the popular Otsu's method in terms of CPU computational time. We demonstrated that our approach outperforms the popular Otsu's method in terms of CPU computational time. We observed a maximum speed-up of 35.58x and a minimum speed-up of 10.21x on popular image processing benchmarks. To demonstrate the correctness of our approach in determining threshold values, we compute PSNR, SSIM, and FSIM values to compare with the values obtained by Otsu's method. This evaluation shows that our approach is comparable and better in many cases as compared to well known Otsu's method.

Motivation & Objective

• To develop a fast and efficient multi-threshold image segmentation technique that avoids the high computational cost of traditional methods like Otsu’s.
• To automatically determine optimal threshold values by identifying valleys in the image histogram, leveraging a heuristic sampling strategy.
• To improve computational efficiency without sacrificing segmentation quality, particularly for multilevel thresholding.
• To provide a scalable and implementable alternative to Otsu’s method, especially for real-time or resource-constrained applications.
• To validate the method’s performance against established metrics (PSNR, SSIM, FSIM) and computational time across diverse image benchmarks.

Proposed method

• The method begins by applying a gradient descent search to identify all local minima (valleys) in the image’s intensity histogram.
• The histogram is partitioned into equal-sized regions to locate the minimum frequency within each, generating candidate threshold values.
• Candidate thresholds are selected to ensure uniform distribution across the histogram, reducing redundancy from multiple local valleys in a single region.
• A naive clustering approach is used to group pixels based on the selected thresholds, with the final threshold values emphasized at valley points.
• The algorithm prioritizes valley points as optimal thresholds, aiming for natural segmentation boundaries between image regions.
• The method is implemented in MATLAB and evaluated using standard image benchmarks, with performance compared to Otsu’s method on both speed and quality metrics.

Experimental results

Research questions

• RQ1 Can a histogram sampling-based approach efficiently and automatically determine multilevel thresholds without relying on computationally expensive optimization?
• RQ2 How does the proposed method compare to Otsu’s method in terms of computational speed across diverse image benchmarks?
• RQ3 To what extent does the segmentation quality of the proposed method match or exceed that of Otsu’s method, as measured by PSNR, SSIM, and FSIM?
• RQ4 Can uniform distribution of candidate thresholds across the histogram improve segmentation stability and reduce overfitting to local noise?
• RQ5 What are the limitations of the current naive clustering approach in threshold selection, and how might advanced clustering improve performance?

Key findings

• The proposed AMTIS method achieves a maximum speed-up of 35.58× and a minimum speed-up of 10.21× over Otsu’s method on standard image benchmarks, demonstrating significant computational efficiency.
• AMTIS outperforms Otsu’s method in terms of CPU computational time across all tested benchmarks, with consistent speed improvements.
• Segmentation results using AMTIS are perceptually clearer and comparable to Otsu’s, with PSNR, SSIM, and FSIM values indicating equivalent or better quality in many cases.
• The method successfully identifies and utilizes histogram valleys as optimal threshold candidates, ensuring natural segmentation boundaries.
• The algorithm occasionally fails to compute optimal thresholds on certain benchmarks due to the limitations of the naive clustering approach, indicating a need for improved clustering heuristics.
• Visual results on benchmarks such as Lena, Cameraman, Baboon, and Finger-Print show that AMTIS produces human-perceptible, high-quality segmentations with clear object-background separation.