Abstract: Hyperbola extraction in ground penetrating radar (GPR) images is an important feature to
analyze the location and structure of underground objects. However, there are often some problems
with the extracted hyperbola, such as incomplete structure, fragmentation, and shape anomalies,
caused by the interference of noise and clutter that are typical of real environments. These issues are
not conducive to the subsequent quantitative operations, such as data analysis and 3D modeling. In
this context, this paper proposed a multi-label hierarchical clustering-based hyperbola extraction
method (MHCE) for the hyperbola extraction of GPR images. Firstly, through evaluating the stability
between pixel neighborhoods by the means of information entropy, an information entropy-based
distance method was constructed to conduct the hierarchical clustering algorithm. Next, a multi-label
clustering method was proposed based on the adjacency space of the clustering results, so as to reduce
the influence of clutter and noise on hyperbola extraction. Finally, the hyperbola was extracted
combined with the fitting shape and texture orientation of the multi-label clustering results. The
experimental results show that this method is robust for GPR images and can be used to obtain the
shape and position parameters of a normalized hyperbola.

Key words: ground penetrating radar image, hyperbola, information entropy, multi-label hierarchical clustering algorithm, robustness