| Abstract: |  This thesis focuses on the problem of automatic identification of tree species based on images of leaves and bark. We propose to describe both leaves and bark using textural features. FSRIT (Fast Scale- and Rotation- Invariant Texture), a novel method for texture description and recognition, is introduced. The method combines an improved scale space (used for multi-scale representation and scale invariant matching) with several state-of- the-art approaches (including LBP-HF features and use of linear SVM classifiers with approximate kernel map). Using the proposed method we achieve new state of the art results in the classification of bark (Austrian Federal Forests bark dataset) and leaves (Austrian Federal Forests leaf dataset, Flavia dataset, Foliage dataset, Swedish dataset and Middle European Woods dataset), as well as on standard textural datasets KTH-TIPS2a and KTH-TIPS2b, while achieving 99% accuracy on all other standard textural datasets (KTH-TIPS, CUReT, UIUCTex, UMD and Brodatz32). The proposed recognition method is very fast and thus suitable for real time applications, including e.g. mobile field guides for plant identification.
|
|---|