Evolutionary optimization of a multiscale descriptor for leaf shape analysis

Abstract

Shape analysis and recognition play an important role in the design of robust and reliable computer vision systems. Such systems rely on feature extraction to provide meaningful information and representation of shapes and images. However, accurate feature extraction is not a trivial task since it may depend on parameter adjustment, application domain and the shape data set itself. Indeed, there is a demand for computational tools to understand and support parameter adjustment and therefore unfold shape description representation, since manual parameter choices may not be suitable for real applications. Our major contribution is the definition of an evolutionary optimization methodology that fully supports parameter adjustment of a multiscale shape descriptor for feature extraction and representation of leaf shapes in a high dimensional space. Here, intelligent evolutionary optimization methods search for parameters that best fit the normalized multiscale bending energy descriptor for leaf shape retrieval and classification. The simulated annealing, differential evolution and particle swarm optimization methods optimize an objective function, which is based on the silhouette measure, to achieve the set of optimal parameters. Our methodology improves leaf shape characterization and recognition due to the intrinsic shape differences which are embedded in the set of optimized parameters. Experiments were conducted on public benchmark data sets with the normalized multiscale bending energy and inner-distance shape context descriptors. The visual exploratory data analysis techniques showed that the proposed methodology minimized the total within-cluster variance and thus, improved the leaf shape clustering. Moreover, supervised and unsupervised classification experiments with plant leaves accomplished high Precision and Recall rates as well as Bulls-eye scores with the optimized parameters.