The statistical analysis of plant part appearance ¯¯ a review

Visual appearance is widely used in assessing the variety or cultivar of crop samples, and in biological taxonomy in general. With computer storage capabilities becoming more easily accessible, image databases of plant specimens can be readily constructed. To make full use of this powerful facility, an important task is to compare different images, or to summarise them. This review addresses the question of how this may be done, given that images are stored as typically 105¯107 highly structured measurements. Image analysis is a wide subject area, and this review covers those parts of it which are relevant to this task. The task is subdivided into four (2×2) categories according to whether we are interested in outline shape only or in colour details, and whether the global position of object details is important, or only their local distribution. We refer to these approaches as local shape, global shape, colour distribution and colour detail (eigenimage analysis). As far as possible, we avoid a priori ideas regarding what features of objects are important, and seek methods which capture the full variability of their appearance. The techniques are illustrated by reference to work on carrots.