Unable to connect to database - 11:36:33 Unable to connect to database - 11:36:33 SQL Statement is null or not a SELECT - 11:36:33 SQL Statement is null or not a DELETE - 11:36:33 Botany & Mycology 2009 - Abstract Search
Unable to connect to database - 11:36:33 Unable to connect to database - 11:36:33 SQL Statement is null or not a SELECT - 11:36:33

Abstract Detail

Systematics Section

Wilf, Peter [1], Chikkerur, Sharat [2], Little, Stefan A. [1], Wing, Scott L. [3], Serre, Thomas [2].

Leaf Identification Automated Using a Computational Model of the Primate Vision System: Preliminary Results.

Angiosperm leaf identification from shape and venation characters is a challenging problem in botany and especially in paleobotany, because the macrofossil record of angiosperms is dominated by isolated leaf impressions. Recognition of diagnostic characters remains inadequate given the complex and variable universe of leaf architecture. Computer algorithms based on primate vision research offer automated training capability and processing power with the potential to discriminate natural groups, even from incomplete specimens. We present preliminary results using cleared-leaf images scanned from the Klucking Leaf Venation Patterns series, digitally stripped of background objects and petioles and checked for APG II status, resulting in five well-sampled families: Combretaceae, Melastomataceae, Passifloraceae, Phyllanthaceae, and Salicaceae (~1400 total images). We used a biologically inspired computer-vision system (Serre et al., PAMI 2007), based on a hierarchy of visual processing stages that closely mimics the primate visual cortex. Shape information (leaf outline and venation) is first processed by an array of local filters at multiple orientations and scales. Features of higher complexity based on combinations of orientations (e.g., junctions between veins, corners, etc.) are then extracted and passed to build a linear classifier for each family. Half the images were randomly chosen to train the system and the remaining half to test it; the procedure was repeated ten times and the reported performance averaged. Tests placed 74.7% +/- 1.7% of leaves in the correct family (chance level 20%). A binary classification between the relatively stereotyped Melastomataceae vs. the rest was correct at 92.7% +/- 1.04% (chance level 50%). These results show high potential for family identification of isolated leaves. More training images for other clades are being assembled, and the general computer model is being adjusted for visual characteristics specific to leaves. We anticipate improved performance on modern leaves, with a goal of testing well-identified fossils.

Log in to add this item to your schedule

1 - Pennsylvania State University, Department of Geosciences, University Park, PA, 16802, USA
2 - Massachusetts Institute of Technology, McGovern Institute for Brain Research, Brain and Cognitive Sciences Department, Cambridge, MA, 02139, USA
3 - Smithsonian Institution, Department of Paleobiology, P.O. Box 37012, NHB MRC 121, Washington, DC, 20013, USA

leaf architecture
computer vision
object recognition
fossil leaves

Presentation Type: Oral Paper:Papers for BSA Sections
Session: 67
Location: Wasatch B/Cliff Lodge - Level C
Date: Wednesday, July 29th, 2009
Time: 1:45 PM
Number: 67004
Abstract ID:622