Chrysanthemum Leaf Classification
The work on Chrysanthemum leaf classification was motivated by contact with the National Institute for Agricultural Botany in Cambridge. Plant breeders who develop a new variety of plant are granted exclusive rights to market that variety for a period of time. One of the requirements of current Plant Breeders Rights legislation is the distinctness of the new varieties: they should be different in at least one character from all existing varieties. The distinctness test is carried out by experts at NIAB which has over 3000 registered varieties of Chrysanthemum, each represented by 10 leaf images. They receive about 300 new applications to be tested each year. NIAB experts utilize a number of heuristic features which are not well defined.
The following are four sample leaves from the Laustin category:
The following are four sample leaves from the Lpalma category:
and the following are four sample leaves from the Lyllwyoda category:
Considering these images, one can appreciate that the problem of automatic classification of leaf images is a difficult task:
We therefore decided to ease the process of test and classification by applying the SQUID system to this problem. SQUID is an image database retrieval system that utilizes shape for similarity retrieval. Our leaf database consisted of 400 leaf images from 40 varieties. Each image was processed to recover the leaf contour which was then represented by the maxima of curvature zero-crossing contours in its Curvature Scale Space image as well as three global shape parameters (eccentricity, circularity, and aspect-ratio of its CSS image). When matching two CSS images, the main idea is to find the optimal horizontal shift that results in the best possible overlap of maxima from the two images. The sum of the Euclidean distances (in CSS) between the corresponding (closest) pairs of maxima is then defined to be the matching value.
- The between-class similarity is considerable while the within-class similarity is not adequate.
- The texture and color of leaves are rather similar.
- The number of classes is quite large, while the number of samples in each class is quite small.
- Overlaps can occur between adjacent parts of leaves. As a result, there can be major differences among boundary contours of leaves in the same category.
Having accepted an input image, the system selects the best 15 similar images from the databse by first applying the global shape parameters to prune the candidates followed by CSS matching. These images are, in general, from different varieties. The best 5 varieties are then selected according to the number of their samples among the retrieved images. The correct variety was among the top 5 choices of the system for over 95% of the inputs. Note that the system uses no expert knowledge (which is not well defined) about Chrysanthemum leaf categories to make its choices. It is intended to be used as an aid by the expert classifier who will then make the final choice.
Further details can be found in published papers (such as Proc. Scale-Space'97 and Proc. International Conference on Visual Information Systems 2000).
F.Mokhtarian@ee.surrey.ac.uk Jan 2004