Automatic diatom identification including life cycle stages for quantitative analysis and water quality assessment
- Autores: Luis Carlos Sánchez Bueno
- Directores de la Tesis: María Gloria Bueno García (dir. tes.), Gabriel Cristobal Pérez (codir. tes.)
- Lectura: En la Universidad de Castilla-La Mancha ( España ) en 2020
- Idioma: español
- Tribunal Calificador de la Tesis: José Manuel Menéndez García (presid.), Jesús Salido Tercero (secret.), Uriel Rodrigo Nava Velazco (voc.)
- Programa de doctorado: Programa de Doctorado en Ciencias y Tecnologías aplicadas a la Ingeniería Industrial por la Universidad de Castilla-La Mancha
- Materias:
- Texto completo no disponible (Saber más ...)
- Resumen
In this dissertation, the automatic identification of diatoms is discussed including the morphological and structural variations during the different stages of their life cycle. The thesis covers the complete workflow from image acquisition to feature extraction and diatom identification.
During the completion of this thesis, an automatic microscope prototype was developed with the aim to provide an affordable solution for automatic diatom identification. Many different components such as motors on all axes, a digital camera and light filters were attached to a low-cost light microscope, so that, it is possible to capture images in different modalities (e.g. dark field and oblique illumination). The motors and illumination are controlled through an open hardware board Arduino Mega. Finally, a powerful GPU based NVIDIA Jetson TX2 board was used to control the microscope and execute image processing and classification tasks efficiently.
The automatic identification method proposed in this work is composed of a segmentation stage followed by the extraction of morphometric and texture descriptors, such as log-Gabor features, phase congruency features and elliptical Fourier Descriptors. Those features were used for a better characterization of the shape of the diatom and to detect texture morphological changes that occur during the diatom's life cycle.
The proposed set of features was assessed on a dataset of 14 diatom species and more than 40 images per taxa. For this dataset different supervised and non-supervised classifiers were used for testing the performance of the selected features. Hierarchical Agglomerative Clustering achieved the best classification results with 99.7% accuracy.
