TMCnet News
Identification and Shape Analysis of Arabidopsis Cultivated in Nitrogen-free Environment [Sensors & Transducers (Canada)](Sensors & Transducers (Canada) Via Acquire Media NewsEdge) Abstract: This paper presents a method for segmentation and shape description of Arabidopsis plants with non-green leaves. The image was first calibrated by detecting the comers of a checkerboard. After the preprocessing step, the image was transformed to CIELUV color space, removing the lightness from the chromatic coordinates. The U component showed markedly different textures between the plant and the background. Hence its standard derivation was calculated and thresholded. With this method, significant leaves of the plant were separated while some stalks were not. Therefore, Support Vector Machine was then used to train the LUV data to do further segmentation as a complement of texture analysis. With these two steps, the plant was completely identified and the shape features were then extracted, including the total area, the symmetry and the number of leaves. The real area of the plant was derived with the number of foreground pixels and the calibration result. The symmetries were represented with the degrees of bilateral symmetry in the direction of the major and minor axes. And the number of leaves was obtained by identifying the number of local maximum of the contour-based signature. Experiment result shows that this method is effective in segmentation and shape analysis of Arabidopsis plants. Copyright © 2014 IFSA Publishing, S. L. Keywords: Arabidopsis, Segmentation, Shape analysis, CIELUV color space, Texture analysis, Support vector machine. (ProQuest: ... denotes formulae omitted.) 1. Introduction Since the genome sequence of Arabidopsis was completed in 2000 [1], a complete understanding of the genes' fun has been anticipated. However, the genetic mechanisms are very complicated. They not only determine the developments of plants, but also react to the environment [2]. In this case, the phenomics research has been thought to be indispensible to realize the goal [3]. With high-throughput analysis of phenotypes, a phenotypic database can be built. The data can be analyzed with the growth and development of plants, genetic variation and environmental interaction, which will contribute to crop cultivation and yield improvement [4]. The phenotypes of Arabidopsis include the length of primary root, the number of leaves, the total leaf area and the perimeter et al., which reflect the growth stage of a plant [5]. PHENOPSIS platform developed an information system called PHENOPSIS DB that shares comprehensive resources for phenotype analysis of Arabidopsis, including a free software platform named ImageJ macros for semi-automatic image analysis. It can be used to analyze the surface area, perimeter, width and height of an Arabidopsis plant [6]. This software, however, works well only when the plants in the image are green. In other words, it is only suitable for the plants which are grown under rigorously controlled environments. In this paper, the research objects are Arabidopsis plants grown under the nitrogen-free condition. Some leaves of the plants turn yellow or other non-green colors, which cannot be detected by this software. Therefore, a novel approach needs be developed to adapt to this research, thus the application can be more flexible and can extract the required features. The first step to analyze a plant is to separate it from the background. Previous researches illustrated many methods to do plant segmentation. Some techniques take advantage of the color features of a pixel [7]. Specifically, the red (R), green (G) and blue (B) components consist a plant pixel and a background one have different ranges of values, respectively. Therefore, their combination can be used to classify a pixel. Woebbecke et al. [8] proposed Excess Green index (ExG) in 1995. Later in 1998, Meyer et al. [9] put forward another index Excess Red (ExR). The combination of these two indices (Eq.l) can be used to segment plants [10]. For a pixel, if the result is larger than 0, it is classified as plant pixel. Some other methods define a decision surface by training Bayes classifier [11] or neural network [12] in color space. Similar to the software ImageJ, these methods are only applied to segment green vegetation. ... (1) where r, g and b are the chromatic coordinates of a pixel. In this paper, the color features of plants are quite complicated. Some leaves are similar to background in color representation. Considering that the plant leaves are smooth, while the background is just on the contrary, of which the pixel values have more changes at local regions, we will combine color features with texture features of the image to do segmentation. As long as the plant is identified, the shape parameters can be calculated, including total area, symmetry of the plant and the number of leaves. The area and the number of leaves show the growth status of a plant, whilst the symmetry reflects morphology and growth orientation. The methods to obtain the data will be proposed and the results will be discussed. 2. Image Acquisition The Arabidopsis plants used were Colombian species. They were mutants with disrupted AMT1.1 gene. The plants were cultivated in vermiculite growth medium and were fertilized with nitrogenfree nutrient solution. At each pot, a small square of checkerboard, with sides of 25 mm, was used to do image calibration. The checkerboard pattem has alternating dark and bright grids, making the grid comers very strong features to be detected [13]. The top views of plants were taken by a camera (NIKON D7000), which the image resolution was set to 1024^678 pixels. One of the original images is shown in Fig. 1(a). 3. Image Preprocessing This part includes comer detection and removal of some background. The first step was completed with the popular method by Zhang [14]. The algorithm is based on Harris comer finder method [15]. This approach achieved good results, as shown in Fig. 1(b). The drawback is that the coordinates of initial guess comers need to be input first. With the comers' coordinates, the image could be corrected. The image correction result is shown in Fig. 1(c). The size of each pixel could be derived with Eq. 2, which could be used to compute the real area of a plant. More details could be referred to [16]. ... (2) where Ac is the real area of each grid of the checkerboard. Nc is the number of pixels of each grid in the image. Pp is the real area of each pixel. To make the further segmentation simpler, the next step was to remove the checkerboard, the pot edge and the orange regions outside the pot. The method was to threshold their standard deviations of their pixels' RGB values. Specifically, for the checkerboard and the pot edge, the RGB values of the pixels are close (all very high for white color and all near 0 for black color), so their standard deviations are smaller than others. The threshold selected was 0.05 from Fig. 2(a). Pixels with standard deviations lower than the value are shown in Fig. 2(b). From the result, the maximum and minimum column coordinates (Xmin, Xmax) and row coordinates (Ymin, Ymax) of these pixels were obtained. Regions outside the range ((Xmin+50): (Xmax-50), (Ymin+50): (Ymax-50) ) were set as 0. Meanwhile, the region with the largest number of pixels was also detected. It is the region of the checkerboard. With closing operation, the holes within it were filled and the region was also removed from the original image completely. With these two steps, the background was composed of only vermiculite medium, making the further plant identification more convenient. 4. Color Space Transform For further processing, the image was transformed from RGB color space to CIELUV (LUV) space. This is due to the fact that in the former space, the three coordinates have strong correlations. In other words, they are all closely related to brightness value. As long as the luminance alters, the three coordinates will all alter. While on the contrary, LUV color space is a perceptually uniform space, where L is for lightness and U and V are chromaticity coordinates [17]. The color space represents luminance by an axis, making the three components independent and less sensitive to the changing of light. To obtain LUV space from RGB coordinates, the values need to be transformed to CIEXYZ (XYZ) space first, as shown in Eq. 3. ... (3) With the XYZ values, the LUV components are computed with Eq. 4. ... (4) where U', V', Un' and V"' are given as: ... (5) where Xn, Yn and Zn are the X, Y and Y values of a specified white point, which are derived with Eq. 3. The results are shown in Fig. 3. 5. Image Segmentation 5.1. Texture Analysis From the U component of the image, it can be seen that the textures of the plant and the background are quite different. The region of plant is smooth, having little change of gray value in a local region, while the background is just the opposite. The standard derivation of the U image is calculated with a 5x5 sliding window. From the histogram of the result shown in Fig. 4(a), the threshold was selected as 0.05. For a pixel, if its standard derivation is smaller than 0.05, as well as the gray value is not 0, it is classified as a plant pixel. The result is shown in Fig. 4(b). It can be seen that most of the leaves are detected. However, some of the stalks are not identified. Besides, the regions of leaves are smaller than the original image. This method could not segment the whole plant region. 5.2. Support Vector Machine Since only texture analysis did not achieve a satisfactory result, the color features were considered to be analyzed as a complement. This step was completed by using Support Vector Machine (SVM). SVM has been widely used for classification. It can find the hyper-plane that maximizes the separating margin as well as minimize the structural risk, making this algorithm advantageous than other traditional methods [18]. The first step is to train the data of the two classes. To be specific, 1000 plant and background points were randomly selected from the LUV components. Then they were labeled and were used to build a model by SVM. The model was then used to predict new examples. The result is shown in Fig. 4(c). Compared with the result of texture analysis, this method could not recognize the non-green leaves. Nevertheless, it reserved the stalks and plant region was not reduced. 5.3. Combination of these Two Methods The two segmentation results were superimposed and binarized with the threshold 0. The final segmentation result is shown in Fig. 5. It was obtained by selecting the region with largest number of pixels and filling the holes in leaves with closing operation. 6. Shape Analysis With the segmentation result, the shape features of the plant could be obtained, including the total area, the symmetry and the number of leaves. 6.1. Total Area As mentioned in preprocessing step, the size of each pixel was obtained. Combined with the number of pixels of the segmentation result, the real area of the plant was derived as Eq. 6. ... (6) where Pp is the size of each pixel. Np is the number of pixels of the plant region in an image. Ap is the real area of the plant. 6.2. Symmetry This feature is described with the degree of bilateral symmetry [19]. There are four steps to compute the value. First, find the major axis and minor axis of the plant. This was done by calculating the eigenvectors of the covariance matrix of the foreground pixels in x and y coordinates. Second, rotate the image to the direction where its major axis is in the horizontal direction, as shown in Fig. 6(a). For this plant, the angle between the major axis and the horizontal direction is 21.8°. Denote A is the new image and N is the number of pixels of the foreground. Then, reflect A with respect to the new major axis (horizontal direction) and minor axis (vertical direction). Thus we obtain two reflected versions B and C. Finally, superimpose B and C onto A, respectively, which are shown in Fig. 6(b) and Fig. 6(c). Let N2b and N2c be the number of pixels whose gray values are 2 in the two superimposed results. The symmetries of major and minor directions are given as N2b / N and N2c / N. 6.3. The Number of Leaves Taking into account that the leaf apex has longer distance than the contour points besides it, the contour - based signature [19] was used to identify the number of leaves. First, the inner holes of the plant region were filled and then the image was blurred so that the outline was smooth, as shown in Fig. 7(a). Second, the contour points were obtained with boundary point tracking algorithm [20]. Then, the distances between the contour points and the centroid of the region were calculated and were stored as a ID signal. Finally, the local maximum points of the signal were selected as the apexes of leaves. For a point, if it is larger than its former point and its next point in the boundary series, it is recognized as the local maximum. From Fig. 7(b), it can be seen that the entire local maximum locations are detected, but some local maximum locations have more than one point due to noise. The problem was solved by using a threshold to process the data. For the points detected, if their locations are within the threshold, only the one with the maximum distance with the centroid was selected as the local maximum. The threshold used was 25 in the experiment. The final result is shown in Fig. 7(c). 7. Results and Conclusions The result of segmentation and measurement is shown in Fig. 8. From the result, it can be easily observed the total areas of plants, their symmetry and the number of leaves. For the Arabidopsis plants represent complex colors when grown in nitrogen-free environment, the texture and color features are combined to do the image segmentation. First, the image was transformed to LUV color space. Second, the standard derivation of the U component was calculated for texture analysis. Then, the LUV data were trained and tested with SVM algorithm. With the two results, the plant was recognized. The real size of the total area of the plant was derived by comparing with a checkerboard served as the reference. The symmetries were represented in the directions of the major and minor axes. And the number of leaves was obtained by identifying the number of local maximum of the contour-based signature. Experiment shows that this method is effective in segmentation and shape analysis of Arabidopsis plants. With the results, it can be easily observed the growth stage of plants. This system is semi-automatic, for it needs to input the coordinates to do comer detection. However, if the height of the camera does not change, the calibration can be performed only once. Thus the processing can be completed automatically. Further improvement will be focused on detection of small leaves to improve the accuracy. Acknowledgements This Paper is supported by the Fundamental Research Funds for the Central Universities (Grant No: YX2013-24) and the Scientific Research Foundation for the Returned Overseas Chinese Scholars, State Education Ministry (2010), which are greatly acknowledged by the authors. References [1] . Arabidopsis Genome Initiative, Analysis of the genome sequence of the flowering plant Arabidopsis thaliana, Nature, Vol. 408, 2000, pp. 796-815. [2] . H. Holtorf, M. C. Guitton, R. Reski, Plant functional genomics, Naturwissenschaften, Vol. 89, Issue 6, 2002, pp. 235-249. [3] . D. Houle, D. R. Govindaraju, S. Omholt, Phenomics: the next challenge, Nature Reviews. Genetics, Vol. 11, Issue 12,2010, pp. 855-866. [4] . D. Edwards, J. Batley, Plant bio informatics: from genome to phenome, Trends in Biotechnology, Vol. 22, Issue 5,2004, pp. 232-237. [5] . D. C. Boyes, A. M. Zayed, R. Ascenzi, A. J. McCaskill, N. E. Hoffman, K. R. Davis, J. Görlach, Growth stage-based phenotypic analysis of Arabidopsis: a model for high throughput functional genomics in plants, The Plant Cell, Vol. 13, Issue 7,2001, pp. 1499-1510. [6] . J. Fahre, M. Dauzat, V. Nègre, N. Wuyts, A. Tireau, E. Gennari, P. Neveu, S. Tisné, C. Massonnet, I. Hummel, C. Granier. PHENOPSIS DB: an information system for Arabidopsis thaliana phenotypic data in an environmental context, BMC Plant Biology, Vol. 11,2011, pp. 77-83. [7] . F. D. Smedt, I. Billauws, T. Goedemé, Neural networks and low-cost optical filters for plant segmentation, International Journal of Computer Information Systems and Industrial Management Applications, Vol. 3,2011, pp. 804-811. [8] . D. M. Woebbecke, G. E. Meyer, K. Von Bargen, D. A. Mortensen, Color indices for weed identification under various soil, residue, and lighting conditions, Transactions of the ASAE, Vol. 38, Issue 1,1995, pp. 259-269. [9] . G. E. Meyer, T. W. Hindman, K. Lakshmi, Machine vision detection parameters for plant species identification, in Proceedings of the Conference on SPIE: Precision Agriculture and Biological Quality, Boston, USA, 1 November 1998, pp. 327-335. [10] . G. E. Meyer, J. Camargo Neto, Verification of color vegetation indices for automated crop imaging applications, Computers and Electronics in Agriculture, Vol. 63,2008, pp. 282-293. [11] . L. F. Tian, D. C. Slaughter, Environmentally adaptive segmentation algorithm for outdoor image segmentation, Computers and Electronics in Agriculture, Vol. 21, 1998, pp. 153-168. [12] . D. S. Shrestha, B. L. Steward, E. Bartlett, Segmentation of plant from background using neural network approach, in Proceedings of the Conference on Artificial Neural Networks in Engineering, St. Louis, USA, 4-6 November 2001, pp. 903-908. [13] . W. Sun, X. Yang, S. Xiao, W. Hu, Robust checkerboard recognition for efficient nonplanar geometry registration in projector-camera systems, in Proceedings of the 5th ACM/IEEE International Workshop on Projector Camera Systems, California, USA, 10 August 2008, pp. 2-8. [14] . Z. Y. Zhang, A flexible new technique for camera calibration, Pattern Analysis and Machine Intelligence, Vol. 22, Issue 11,2000, pp. 1330-1334. [15] . C. Harris, M. Stephens, A combined comer and edge detector, in Proceedings of the Conference on Alvey Vision, Manchester, USA, 31 August - 2 September 1988, pp. 147-152. [16] . Q. Ke, J. Zhang, Y. Tian, Arabidopsis growth analysis using image processing technology, Advanced Materials Research, Vol. 712, 2013, pp. 2359-2363. [17] . J. Schanda, Colorimetry: understanding the CEE system, Willey, 2007. [18] . V. Vapnic, The nature of statistical learning theory, Springer, 2000. [19] . L. F. D. Costa, Jr. R. M. Cesar, Shape analysis and classification: theory and practice, CRC Press, 2000. [20] . R. C. Gonzalez, R. E. Woods, Digital image processing, Publishing House of Electronics Industry, 2011. * Junmei ZHANG, Ye TIAN, Qiuhong KE School of Technology, Beijing Forestry University, Beijing, 100083, China * Tel.:+86-1062336398 * E-mail : j oyzhangj m@ 163. com Received: /Accepted: 30 June 2014 /Published: 31 July 2014 (c) 2014 IFSA Publishing, S.L. |