Abstract
Intensity inhomogeneities in images constitute a considerable challenge in image segmentation. In this paper we propose a novel biconvex variational model to tackle this task. We combine a total variation approach for multi class segmentation with a multiplicative model to handle the inhomogeneities. Our method assumes that the image intensity is the product of a smoothly varying part and a component which resembles important image structures such as edges. Therefore, we penalize in addition to the total variation of the label assignment matrix a quadratic difference term to cope with the smoothly varying factor. A critical point of our biconvex functional is computed by a modified proximal alternating linearized minimization method (PALM). We show that the assumptions for the convergence of the algorithm are fulfilled by our model. Various numerical examples demonstrate the very good performance of our method. Particular attention is paid to the segmentation of 3D FIB tomographical images which was indeed the motivation of our work.
Introduction
In image processing, an ideal scenario involves pixels of the same material possessing identical intensity values. However, real-world data often suffers from "intensity inhomogeneities" or bias fields—uneven lighting or sensor artifacts that cause the same tissue or material to appear with varying brightness across an image. Standard segmentation techniques like K-means or global thresholding frequently fail under these conditions because they rely on global intensity consistency.
This paper introduces a novel unsupervised variational model designed to simultaneously segment 3D images and correct for these inhomogeneities. The primary motivation is 3D Focused Ion Beam (FIB) tomography, where lighting gradients pose a significant hurdle in identifying microscopic structures in materials science.
Methodology
The proposed approach utilizes a biconvex variational model. The core premise treats the observed image $F$ as a product of structural reflectance $R$ and a smoothly varying illumination component $L$: $F = R \cdot L$. To simplify the optimization, the model operates in the logarithmic domain ($f = r + l$), transforming the multiplicative relationship into an additive one.
The objective function incorporates three critical components:
- Data Term: A fuzzy membership function assigning pixels to one of $K$ classes based on their fit to the corrected intensity.
- Total Variation (TV) Regularization: Applied to the label assignment matrix to ensure smooth region boundaries and noise robustness.
- Smoothness Penalty: A quadratic difference term applied to the illumination component $l$ to prevent it from capturing sharp structural details.
Optimization is achieved through a Modified Proximal Alternating Linearized Minimization (PALM) algorithm. This method iteratively updates segmentation labels, class centers, and the illumination field, ensuring convergence to a stable critical point.
Experimental Results
The model was evaluated on synthetic data, medical imagery (X-ray and CT angiography), and 3D FIB tomography. In synthetic tests, the method demonstrated superior noise robustness compared to state-of-the-art level-set methods, maintaining low error rates even under high Gaussian noise. In medical contexts, it successfully preserved thin structures, such as blood vessels, which are often lost during traditional two-step bias correction.
The primary success was observed in 3D FIB tomography of an aluminum matrix containing silicon carbide (SiC) particles and copper aggregations. Despite significant intensity gradients across the 3D volume, the model accurately identified all three phases, effectively handling the diverse shapes and sizes of the particles.
Discussion and Conclusion
A key insight of this work is the advantage of simultaneous bias correction and segmentation. By coupling these tasks, the segmentation informs the illumination estimate about structural edges, while the illumination estimate ensures segmentation accuracy across the volume. The biconvex nature of the problem allows for efficient computation using the PALM solver, making it suitable for high-dimensional 3D data.
Future work may involve enforcing hard binary constraints on labels to improve edge sharpness, extending the model to multi-channel (color) data, and incorporating deconvolution operators to address optical blurring.
Summary updated