Hierarchical DNN-Based Image Segmentation Algorithm Using Texton, Superpixels, and Layer-Adaptive Loss Functions
Journal
Proceedings - 16th International Conference on Signal-Image Technology and Internet-Based Systems, SITIS 2022
ISBN
9781665464956
Date Issued
2022-01-01
Author(s)
Yu, Cheng Hsuan
Abstract
Image segmentation is critical to object-oriented image processing. Many conventional segmentation algorithms are based on the superpixel, since it integrates the pixels with similar colors and locations in prior and is beneficial for segmentation. Recently, several segmentation algorithms based on deep learning were developed. However, due to the irregular shape and size of superpixels, it is hard to apply the superpixel directly in a leaning-based segmentation algorithm. In this paper, we propose a novel segmentation method that well integrates the techniques of the deep neural network (DNN), the superpixel, adaptive loss functions, and multi-layer feature extraction. First, different from other learning-based algorithm, which applies an image or its bounding boxes as the input, we adopt the mean and the histogram differences of the features of two superpixels as the input of the DNN to determine whether they should be merged. Moreover, to well consider both large- scaled and small-scaled features, a hierarchical architecture is adopted. For different layers, the DNN models with different loss functions are applied. A larger penalty for over-merging is applied in the first layer and a larger penalty for over-segmentation is applied in the following layer. Moreover, according to human perception, the features of colors, areas, the gradient at the boundary, and the texton, which is highly related to the texture, are applied. Experiments show that the proposed method outperforms other state-of-the-art image segmentation methods and produces highly accurate segmentation results.
Subjects
deep learning | image segmentation | superpixel merging
SDGs
Type
conference paper