LEAF SEGMENTATION AND PARALLEL PHENOTYPING FOR THE ANALYSIS OF GENE NETWORKS IN PLANTS Olivier Janssens 1,2 , Jonas De Vylder 3 , Jan Aelterman 3 , Steven Verstockt 2 Wilfried Philips 3 , Dominique Van Der Straeten 4 , Sofie Van Hoecke 1,2 , Rik Van de Walle 2 1 Electronics and Information Technology lab, ISP, Ghent University Graaf Karel de Goedelaan 5, 8500 Kortrijk, Belgium Telephone: + 32 56 24 12 52; email: odjansse.janssens@ugent.be 2 Multimedia Lab, ELIS, Ghent University – iMinds 3 Dept. of Telecommunications and Information Processing, TELIN, Ghent University – iMinds 4 Laboratory of Functional Plant Biology, Department of Physiology, Ghent University ABSTRACT Over the last 4 years phenotyping is becoming more and more automated, decreasing a lot of manual labour. Features, which uniquely define the plant, can be extracted automatically from images. As a lot of plant data has to be processed in order to extract the features, fast processing of these features is a chal- lenge. Therefore in this paper, a new method for automatic segmentation of individual leaves from plants with a circular arrangement of leaves (rosettes) is proposed, together with an algorithm to extract the line of symmetry of the leaf. Fur- thermore, in order to achieve fast processing for phenotyping plants, four feature extraction methods are parallelised in or- der to run on the CPU and GPU. Our evaluation results show that by parallelizing the feature extraction methods, it is possi- ble to calculate the image moments, area, histogram and sum of intensities 5 to 45 times faster than single threaded imple- mentations. Index Terms— segmentation, parallelisation, phenotyp- ing, OpenCl, image processing 1. INTRODUCTION Plant analysis by computer vision is a promising non-destruc- tive method which can be performed in an automatic way. Several methods have been proposed in literature which fo- cus on the leaves solely, such as the LIMANI [1] framework or LEAF GUI [2]. The LIMANI framework focuses on the automatic segmentation and measurement of venation pat- terns. LEAF GUI is a user-assisted software tool that facili- tates improved empirical understanding of leaf network struc- ture. These frameworks are of value for analysing plants, though the disadvantage is the fact that leaves are not seg- mented automatically. Apart from the venation patterns it is also possible to extract other features such as plant growth [3], area of the leaves, relative growth rate, compactness of the rosette, diameter of the rosette, stockiness, and intensity of the plant image [4–7]. The biggest limitation, however, is that these features are usually not extracted from the leaves solely, but rather from the entire rosette. All these features have to be calculated for each plant and for every time frame, requiring a lot of processing time, which can introduce a bottleneck in the analysis of plants. Because of the before mentioned problems, we can state there is a need for (1) automatic leaf segmenta- tion (i.e., in the discussed state-of-the-art leaf segmentation is done manually) and (2) fast processing for feature extraction to cope with growing datasets. Therefore, an algorithm capa- ble of segmenting individual leaves is presented in this paper. Additionally, new and generic features are proposed that al- low for parallel processing in order to achieve fast processing of the plants. The remainder of this paper is as follows: Section 2 ex- plains how automatic leaf segmentation from a plant rosette can be done. Subsequently section 3 introduces the algorithm to extract the line of symmetry from a leaf. Next, Section 4 gives an overview of features that are extracted in parallel from the rosette. Evaluation results are presented in Section 5. Finally, Section 6 ends this paper with conclusions. 2. LEAF SEGMENTATION In order to be able to extract features accurately, it is nec- essary to segment the rosette as precise as possible. As de- scribed by De Vylder et al. [7], several effective and effi- cient methods exists to extract individual rosettes such as su- pervised pixel classification methods, pixel clustering based EUSIPCO 2013 1569744447 1