SuperBF: Superpixel-Based Bilateral Filtering Algorithm and Its Application in Feature Extraction of Hyperspectral Images

Zhikun Chen,Zhihua Cai,Chong Zhou,Shaoyuan Fu,Junjun Jiang

doi:10.1109/access.2019.2938397

Abstract

Bilateral filtering (BF), which is an edge-preserving filtering (EPF) method, has been widely recognized as a simple and efficient approach for hyperspectral image (HSI) feature extraction. However, due to the limitation of spatial resolution and the influence of the complexity of land feature distribution in HSIs, updating the target pixel through weighted averaging of neighbouring pixels is prone to generating mixed pixels, i.e., the updated target pixel is mixed with the feature of other land objects in addition to that of the target object, decreasing the quality of the image feature extraction. To address this problem, in this study, we propose a superpixel-based BF algorithm, SuperBF. This algorithm divides a HSI into many homogeneous regions via superpixel segmentation and then separately filters each homogenous region via BF; this approach ensures that the pixel structure in the template after BF is similar to that in the filtering process, reduces the probability of generating mixed pixels, and thus improves the quality of the image feature extraction. To verify the effectiveness of this proposed method, a support vector machine (SVM) classifier is used to classify the extracted SuperBF features. Experiments on three commonly employed HSI datasets demonstrated that SuperBF is significantly superior to the traditional BF-based hyperspectral feature extraction method and some new feature extraction methods.

Highlights

Ahyperspectral image is a digital image of hundreds of narrow spectral bands and visible infrared spectral bands acquired by satellite sensors [1]–[4]
A superpixel BF algorithm (SuperBF) algorithm was proposed in this study; the algorithm reasonably divides an image into homogeneous regions via superpixel segmentation
The support vector machine (SVM) algorithm was implemented in the libsvm [39] library with five-fold cross-validation, and the default parameters in the references were employed in other algorithms

Summary

INTRODUCTION

Ahyperspectral image is a digital image of hundreds of narrow spectral bands and visible infrared spectral bands acquired by satellite sensors [1]–[4]. Pan et al [27] proposed the R-VCANet deep learning method, which can combine spectral and spatial features using a rolling guide filter (RGF) and extract the depth features of HSIs using the new vertex component analysis network (VCANet). The second section briefly introduces the entropy rate superpixel segmentation (ERS) algorithm and the related topic of BF; it describes the feature extraction algorithm for HSIs based on SuperBF. The segmentation result is obtained by iteratively maximizing this objective function This method projects the image to an undirected graph G = (V , E), where V is the set of vertices of the graph, E is the set of edges of the graph, and the weights of the edges represent the similarity among the vertices, which is quantified by the weight function ω : E → R+ ∪ {0}.

SUPERBF-BASED FEATURE EXTRACTION ALGORITHM FOR HSIS

Result

EXPERIMENTAL RESULTS

CONCLUSION AND FUTURE WORK