This paper deals with the analysis of images of biological tissue that involves ring structures, such as tree trunks, bivalve seashells, or fish otoliths, with a view to automating the acquisition of age and growth data. A bottom-up template-based scheme extracts meaningful ridge and valley curve data using growth-adapted time-frequency filtering. Age and growth estimation is then stated as the Bayesian selection of a subset of ring curves, which combines a measure of curve significativity and an a priori statistical growth model. Experiments on real samples demonstrate the efficiency of the proposed data extraction stage. Our Bayesian framework is shown to significantly outperform previous methods for the interpretation of a data set of 200 plaice otoliths and compares favorably with interexpert agreement rates (88% of agreement to expert interpretations).
Problem statement and related workAge and growth data provide key information for a broad range of scientific issues: for instance, the computation of daily temperature series for paleoclimatology from the analysis of daily increments 30 on the shell of sea scallops (Chauvaud 2004), the determination of age-length keys from the interpretation of fish otoliths for marine stock assessments (Panfili et al. 2003), the analysis of fish otolith growth patterns for marine ecology (Hare and Cowen 1997;Panfili et al. 2003) or tree-ring dating for archeology (Baillie 1982). All of these applications initially rely on the interpretation of growth rings observed on biological materials such as tree trunks, corals, shells of seashells, fish otoliths, fish scales, 35 which will be subsequently referred as biological hard tissues (Fig. 1). The presence of concentric ring structures is due to the variations in chemistry during the accretionary formation of these structures.The deposit of the rings is often associated with a biological periodicity (either seasonal or daily, depending on the species and the tissue). Counting growth rings then leads to an estimation of the age of the analyzed tissue, whereas growth increments can be estimated from the distance between successive 40 rings (Fig. 2), however, not all the observed ring structures are relevant for aging (Campana 2001).The previously mentioned biological applications require large collections of age and growth data for a consistent analysis. The acquisition of this data is generally performed by expert readers (typically, for marine stock assessment issues, several thousands of fish otoliths per year). The development of digital imaging systems and computer-aided tools brings new solutions to this field, (Small and 45 Hirschhorn 1987;Guay 1997;Morison et al. 1998). Our work focuses on the automation of age and growth estimation from the analysis of ring structures within the images of biological hard tissues. As detailed below, the key issues lie in the use of growth information to design a ring extraction framework that takes into account growth non-linearity (Fig. 2.b), and to discriminate actual growth rings...