José Antonio Soria Pérez
Ph.D. Thesis title:
On the Automatic Detection of Otolith Features for Fish Species Identification and their Age Estimation
Author:
José Antonio Soria Pérez
Director:
Vicenç Parisi Baradad
Reading date:
21/01/2013
Abstract:
This thesis deals with the automatic detection of features in signals, either extracted from photographs or captured by means of electronic sensors, and its possible application in the detection of morphological structures in fish otoliths so as to identify species and estimate their age at death. From a more biological perspective, otoliths, which are calcified structures located in the auditory system of all teleostean fish, constitute one of the main elements employed in the study and management of marine ecology. In this sense, the application of Fourier descriptors to otolith images, combined with component analysis, is habitually a first and a key step towards characterizing their morphology and identifying fish species. However, some of the main limitations arise from the poor interpretation that can be obtained with this representation and the use that is made of the coefficients, as generally they are selected manually for classification purposes, both in quantity and representativity. The automatic detection of irregularities in signals, and their interpretation, was first addressed in the so-called Best-Basis paradigm. In this sense, Saito's Local discriminant Bases algorithm (LDB) uses the Discrete Wavelet Packet Transform (DWPT) as the main descriptive tool for positioning the irregularities in the time-frequency space, and an energy-based discriminant measure to guide the automatic search of relevant features in this domain. Current density-based proposals have tried to overcome the limitations of the energy-based functions with relatively little success. However, other measure strategies more consistent with the true classification capability, and which can provide generalization while reducing the dimensionality of features, are yet to be developed. The proposal of this work focuses on a new framework for one-dimensional signals. An important conclusion extracted therein is that such generalization involves a mesure system of bounded values representing the density where no class overlaps. This determines severely the selection of features and the vector size that is needed for proper class identification, which must be implemented not only based on global discriminant values but also on the complementary information regarding the provision of samples in the domain. The new tools have been used in the biological study of different hake species, yielding good classification results. However, a major contribution lies on the further interpretation of features the tool performs, including the structure of irregularities, time-frequency position, extension support and degree of importance, which is highlighted automatically on the same images or signals. As for aging applications, a new demodulation strategy for compensating the nonlinear growth effect on the intensity profile has been developed. Although the method is, in principle, able to adapt automatically to the specific growth of individual specimens, preliminary results with LDB-based techniques suggest to study the effect of lighting conditions on the otoliths in order to design more reliable techniques for reducing image contrast variation. In the meantime, a new theoretic framework for otolith-based fish age estimation has been presented. This theory suggests that if the true fish growth curve is known, the regular periodicity of age structures in the demodulated profile is related to the radial length the original intensity profile is extracted from. Therefore, if this periodicity can be measured, it is possible to infer the exact fish age omitting feature extractors and classifiers. This could have important implications in the use of computational resources anc current aging approaches.
On the Automatic Detection of Otolith Features for Fish Species Identification and their Age Estimation
Author:
José Antonio Soria Pérez
Director:
Vicenç Parisi Baradad
Reading date:
21/01/2013
Abstract:
This thesis deals with the automatic detection of features in signals, either extracted from photographs or captured by means of electronic sensors, and its possible application in the detection of morphological structures in fish otoliths so as to identify species and estimate their age at death. From a more biological perspective, otoliths, which are calcified structures located in the auditory system of all teleostean fish, constitute one of the main elements employed in the study and management of marine ecology. In this sense, the application of Fourier descriptors to otolith images, combined with component analysis, is habitually a first and a key step towards characterizing their morphology and identifying fish species. However, some of the main limitations arise from the poor interpretation that can be obtained with this representation and the use that is made of the coefficients, as generally they are selected manually for classification purposes, both in quantity and representativity. The automatic detection of irregularities in signals, and their interpretation, was first addressed in the so-called Best-Basis paradigm. In this sense, Saito's Local discriminant Bases algorithm (LDB) uses the Discrete Wavelet Packet Transform (DWPT) as the main descriptive tool for positioning the irregularities in the time-frequency space, and an energy-based discriminant measure to guide the automatic search of relevant features in this domain. Current density-based proposals have tried to overcome the limitations of the energy-based functions with relatively little success. However, other measure strategies more consistent with the true classification capability, and which can provide generalization while reducing the dimensionality of features, are yet to be developed. The proposal of this work focuses on a new framework for one-dimensional signals. An important conclusion extracted therein is that such generalization involves a mesure system of bounded values representing the density where no class overlaps. This determines severely the selection of features and the vector size that is needed for proper class identification, which must be implemented not only based on global discriminant values but also on the complementary information regarding the provision of samples in the domain. The new tools have been used in the biological study of different hake species, yielding good classification results. However, a major contribution lies on the further interpretation of features the tool performs, including the structure of irregularities, time-frequency position, extension support and degree of importance, which is highlighted automatically on the same images or signals. As for aging applications, a new demodulation strategy for compensating the nonlinear growth effect on the intensity profile has been developed. Although the method is, in principle, able to adapt automatically to the specific growth of individual specimens, preliminary results with LDB-based techniques suggest to study the effect of lighting conditions on the otoliths in order to design more reliable techniques for reducing image contrast variation. In the meantime, a new theoretic framework for otolith-based fish age estimation has been presented. This theory suggests that if the true fish growth curve is known, the regular periodicity of age structures in the demodulated profile is related to the radial length the original intensity profile is extracted from. Therefore, if this periodicity can be measured, it is possible to infer the exact fish age omitting feature extractors and classifiers. This could have important implications in the use of computational resources anc current aging approaches.
Share: