Plant identification using local invariants: dense sift approach

Official URL: http://risc01.sabanciuniv.edu/record=b1615028 (Table of Contents)

In this thesis, we investigate the use of Dense SIFT approach in automatic identification of plants from photographs. We concentrate on owering plants and evaluate three alternative approaches. In the first one, we classify the plant directly using the dense SIFT method, using appropriate parameters that are found using experimental validation techniques. In the second approach, we first identify the dominant colour in the photograph and use a separate classifier in each of the colour cluster. The second approach is intended to reduce the problem complexity and the number of classes handled by each classifier. In this approach, the classifier for red owers will not know about a plant that does not ower in red; furthermore a plant that is only observed with red owers will only be handled by that classifier. In a third approach, we precede the second approach by adding a Region of Interest detector, in order to extract the flower color more reliably. We find that enhancement of Dense SIFT features based identification is possible with saturation-weighted hue histogram based color clustering and region of interest detector. Using the proposed system, we obtain a 0:60 accuracy on the ower subset in the LifecLEF 2014 database.