AI Communications 29 (2016) 687–699 687 DOI 10.3233/AIC-160715 IOS Press A rotation-invariant feature space according to environmental applications needs in a data mining system using fish otoliths Pere Marti-Puig ∗ and Ramon Reig-Bolano Data and Signal Processing Group, Department of Digital Information Technologies, University of Vic – Central University of Catalonia (UVic-UCC), 08500 Vic, Barcelona, Spain E-mails: pere.marti@uvic.cat, ramon.reig@uvic.cat Abstract. Otoliths are microcalcifications found in the inner ear of fishes and their shape can be analyzed to determine sex, age, populations and species, and thus they can provide necessary and relevant information for ecological and environmental studies. This paper compares two different approaches to deal with an arbitrary rotation of a query image in a data mining system that classifies fish otoliths from their images according to the needs of a real environmental application. The different approaches proposed in this paper will allow the successful use of images that are not normalized or are normalized in a different positioning than the images in the database, something that could be very useful in real field applications. The first approach tested is based on a Rotation-invariant Feature Space derived from the Elliptical Fourier Descriptors (EFD), and the second approach relies on a rotation estimation module; in both cases the data mining system is based in a multiclass classifier implemented in this test case with a simple k-Nearest Neighbours after a complex pre-processing step that provides the Feature extraction. Keywords: Rotation invariant features, feature space, k-NN classifiers, multiclass classifiers, Elliptic Fourier Descriptor (EFD), pattern matching, pre-processing, data mining, otoliths 1. Introduction Otoliths are calcareous structures found in the inner ear of Osteichthyan fishes. In ichthyology, the charac- teristics of otoliths (size, morphologic specificity, ac- cessibility, chemical composition, microstructure and mode of growth) make them one of the most use- ful anatomic structures for various studies. Moreover, these otoliths properties depend on the variation in en- vironmental and genetic factors [3], which leads to a large number of practical applications [2,9]. These ap- plications are not limited to ichthyology, but are widely extended from the study of the feeding ecology of fish predators (the otoliths found in their stomach pro- vide reliable information on their diet) to some aspects of palaeontology, stratigraphy, archaeology (as these structures can be fossilized) or even zoogeography. An online otolith database named AFORO has been in development by the AFORO Team project since 2005 (http://aforo.cmima.csic.es/)[10]. AFORO is or- ganized into high-resolution fish otolith images sets * Corresponding author. E-mail: pere.marti@uvic.cat. with complete morphometric information and includes a shape analysis module that provides mathematical descriptors of the otoliths [13], like Fourier Transform (FT), Curvature Scale Space (CSS) and Wavelet Trans- form (WT). Later, Parisi et al. [14] developed a data mining system that is able to classify external otolith images comparing query images against the AFORO database. There are several otoliths in the ear but the sagitta is the otolith with the largest morphological variability, and therefore it is the most studied and the one selected for the AFORO database. Throughout the text we will usually refer to any sagitta otolith as “otolith” for simplicity. In this paper we focus on validating an improvement that could be added to the present data mining system modifying the pre-processing part. One of the main drawbacks of the system in its current form is its ex- treme sensitivity to the positioning of the otolith in the query image, especially when dealing with rotations and misalignment of the images respect to the coun- terparts at the database. We will show that the system decreases drastically in performance when the images 0921-7126/16/$35.00 © 2016 – IOS Press and the authors. All rights reserved