Weak-prior medical image matting based on microscale-searching evolutionary optimization

Abstract

Image matting is to predict alpha mattes that reflects the opacity of images, recently showing potential in identifying transition regions of lesions in computer aided diagnosis. Image matting can be modeled as a large-scale combinatorial optimization problem that has numerous subproblems. Evolutionary algorithms (EAs) have been applied to predict accurate alpha mattes. The advantage of EAs-based methods is the ability to predict alpha mattes with weak prior like trimaps that provide value of opacity for pixels effortless to annotate compared to recent deep learning-based methods. However, it is challenging for EAs to solve the problem efficiently due to numerous subproblems and the large size of the decision set. Based on the observation that the similarity of subproblems correlates with the similarity of their objective spaces, this paper proposes a method for estimating a microscale subset of the decision set from the solving process of similar subproblems. A framework is designed to reduce the exploration cost of EAs in the large-scale decision set by guiding EAs to search in this estimated microscale subsets. Three medical image matting datasets are used to validate our method’s improvement in the efficiency of evolutionary algorithms. Experimental results demonstrate that EAs embedded in the proposed framework obtain the best prediction of alpha mattes on medical images and also in weak scenarios involving natural images. Comparative experimental results on multi-objective performance metrics indicate that our method is capable of finding superior solutions using fewer fitness evaluations. The contribution of our work is to make EAs an efficient approach to solving the medical image matting problem with weak prior.