In this paper, we studied the problem of recognizing African clawed
frogs' identities from their skin patterns. We proposed a novel, and as
far as our literature search is accurate, the only frog body localization
and skin pattern region extraction algorithmreported in the open liter-ature. We tested and reported its accuracy and stability, which showed
that it is quite successful in compensating scale and rotation changes.
Moreover, our qualitative observations showed that it extracts majorly
overlapping windows even in the case of intensive modifications.
The pattern of the frog skin is a unique identifier which is used for
manual identification by the experts. Our aim was tofind the best
descriptor to represent the skin patterns. In addition we tried to
seek an answer to the question on whether we needed an abstract
representation at all. Hence, we have comparedfive candidate fea-tures (Gabor filters, area granulometry, HoG, dense SIFT, and raw
pixel values). The detailed experiments using a nearest neighbor
classifier showed that the raw pixel values of the extracted pattern
window (in coarse resolutions) was the most effective feature,
thanks to the stability of the pattern window extractor. However, it
performed the worst against the intensity modification test. In addi-tion, it is not as robust as the SIFT against affine modifications, which
can certainly result in poorer recognition performance during a real
use scenario.
Itmay be possible to create a joint feature using the raw pixel values
andone of the other features, forwhich theHoG, and SIFT features seem
to be good candidates. An important futurework is to test our algorithm
and features on a large database of African clawed frogs which include
pictures taken at different times with varying conditions.
The focus of this study is on the automatic recognition of African
clawed frogs in a laboratory environment. However, we believe that
this work can be extended to the natural habitats by an additional
pre-frog detection step which must locate frogs in uncontrolled envi-ronments. The generalized visual object detectors can be used for this
purpose. Subsequently, our technique for locating the skin pattern
region window can be used. Since it works for frog bodies which are
veryflexible, it can be adapted to other skin-textured or fur-textured
species if adapted to the respective body shapes and geometries. More-over, the comparison of various features presented here must be rele-vant to the different skin–fur textures.