University of Maryland

[CASCI Talk] Automatic Image Annotation Applied to Habitat Classification

November 10th, 2014 by

Nov. 11
2 pm – 3 pm
2116 Hornbake Bldg, South Wing

Mercedes Torres, PhD  
As a post-­‐doctoral student at the University of Nottingham, she is focused on interdisciplinary research, specifically in the area of Fine-­‐Grained Visual Categorization, Image Processing and Analysis, and Machine Learning. She has designed and developed an image annotation framework for Phase 1 habitat classification in ground-­‐taken photographs.

Currently, habitat classification (the process of mapping an area with the habitats present on it) is carried out by human surveyors.This is expensive, time consuming, laborious and subjective. What I have done is develop the first complete automatic alternative for the Phase 1 classification Scheme, widely used in the UK. The problems itself is quite complicated, giving the semantic similarities between the classes I have to recognize. I have approached habitat classification as an image annotation problem and created a complete framework for it, composed of 5 elements: the source data, features extracted from these data (low and medium), a novel machine learning classifier called Random Projection Forests and a location-­‐ based voting system for my classifier. Moreover, I have used a novel source of information as the input of this framework: ground-­‐taken geo-­‐referenced photographs (which can be photographs taken with a mobile phone). Current state-­‐of-­‐the-­‐art normally uses remote-­‐sensed imagery, but this is not detailed enough to distinguish between vegetation species, etc.,so ground-­‐taken photos are actually better alternatives. Additionally, I have created a new ensemble classifier, called Random Projection Forests and based on Random Forests, but much more efficient and accurate. Results show that my complete framework can successfully classify 7 out of the 10 main classes of Phase 1, which is quite good considering that this type of work has never been done before with the type of data I am using and the approach I have chosen.