Rozpoznávání Rostlin a Hub z Obrázků

• Main Author: Šulc, Milan
• Format: Dissertation
• Language: Czech
• Published: ProQuest Dissertations & Theses 01-01-2020
• Subjects: Artificial intelligence; Biodiversity; Botany; Computer science; Conservation biology; Flowers & plants; Fungi; Machine learning; Neural networks; Support vector machines; Taxonomy
• ISBN: 9798802731239

The thesis contributes to fine-grained recognition of plant and fungi species from images, ranging from scans and photos of leaves and bark taken in controlled conditions to unconstrained observations of plants and fungi “in the wild” with complex background and clutter in the scene. The constrained tasks of bark and leaf identification are approached as a texture recognition problem. For more complex species recognition tasks with large scale datasets available, we take a deep learning approach. In many instances of the species recognition problem, test-time categorical priors differ from the training set. We address the problems of adjusting outputs of probabilistic classifiers to new priors and estimating the new priors. In particular, we note that training a neural network by cross entropy minimization leads to a model whose outputs should be an estimate of the posterior probabilities. We experimentally validate related statistical properties of the outputs of Convolutional Neural Network (CNN) classifiers. For estimation of test-time categorical priors, a Maximum Likelihood estimation approach is compared with a proposed Maximum a Posteriori estimation, adding a hyper-prior favouring dense prior distributions. We show that adding such hyper-prior increases the reliability of the estimate and increases the classification accuracy in several fine-grained classification tasks.The proposed texture recognition method, Fast Features Invariant to Rotation and Scale of Texture (Ffirst), achieved excellent results in leaf and bark classification, as well as in standard texture classification. The deep learning approach presented in the thesis has scored first in several species recognition competitions on “in the wild” plant and fungi identification, where the views of the observed specimen vary significantly and the difficulty is increased by occlusions and background clutter. The results confirm the benefits of practices such as combining predictions from an ensemble of models, filtering potentially noisy data, data augmentation, and using the moving averages of the trained variables. An experimental comparison with human experts in plant identification shows that the best ensembles of deep CNNs reach the human expert accuracy in image-based plant identification. The competition-winning model for fungi recognition is applied in a citizenscience project and assists the collection of fungi observations, valuable for several research fields including mycology and biodiversity research.