2010 International Conference on Educational and Network Technology (ICENT 2010)

Anisotropic Diffusion Based Weed Classifier

Shujaat Ali Khan

Institute of Management Sciences Peshawar, Pakistan

shujaat446@yahoo.com

Owais Adnan

Institute of Management Sciences Peshawar, Pakistan

Abstract- This paper presents a new approach of anisotropic diffusion to classify the weed images into broad and narrow

class for real time selective herbicide application The classifier we proposed based on Perona and Malik equation. Its low computational complexity and fast runtimes makes this method well suited for real-time vision applications. The developed system has been tested on weeds in the lab; the

results show a very reliable performance and drastically less computational effort on images of weeds taken under varying

field conditions. The analysis of the results shows over 97.6% classification accuracy over 200 sample images.

Keywords- Image Processing; Anisotropic Diffusion; Real­Time Recognition; Ecology; Weed detection

I. INTRODUCTION

Weed control is a critical farm operation and can significantly affect crop yield. Herbicides have vital importance in weed control and high crop yield however these have potential to produce harmful effects [17]. Herbicides are applied to whole field uniformly without considering the weed density . Weeds are often patchy rather than even or randomly distributed in the crop fields [18]. Total variable costs in 2002 for U.K were within a range of £1,720lha and £1,870lha for main crop potatoes, of which herbicides accounted for between 3% and 4% of costs fungicides accounted for about 8% of variable costs and nematicides accounted for about 14%-16% of variable costs. United States farmers applied about $16 billion of herbicides in 2005 (The Value of Herbicides in U.S. Crop Production: 2005 Update, Crop Life Foundation), in 1965 pesticide use was $474.1 million for the United States. By 1970 the use of pesticides doubled to $960 million for the United States and between 1975 and 1999 pesticide use grew 383% for the United States (Agribusiness and Applied Economics Report No. 456), representing a significant portion of the variable costs of agricultural production.

The amount of herbicides in a control patch sprayer has been potentially reduced when real-time weed sensing is used. Patch spraying using remote sensing and machine vision are successful systems [19].

The objective of this research was to develop a vision algorithm. It was not an autonomous robot but a real-time machine vision system, which can recognize the absence of weed and differentiate the presence of broad leaf weed and

978- 1-4244-7662-6/$26.00 © 20 10 IEEE 1 1

Abdul Muhamin Naeem

ITC&S, NetSol Technologies Peshawar, Pakistan

abdulmuhamin@yahoo.com

Attaullah Khan

Provincial Assembly Secretariat, NWFP Peshawar, Pakistan

narrow leaf weed and also to construct and evaluate a classifier that was capable of recognizing the presence and type of weeds and then the appropriate herbicides could be applied using the automatic sprayer control system based on the proposed algorithm.

Much research has investigated strategies to control weeds with less herbicide to reduce production costs and to environmental pollution.

Although weeds are not evenly distributed on field, weed treatments are mostly applied with the same dose over the entire field, while the herbicides are used more efficiently by adjusting the herbicide dose to specific weed [1]

A verity of visual characteristics that have been used in plant identification can be divided into three categories: Spectral Reflectance, Morphology and texture. The photosensor-based plant detection systems [2], [3] can detect all the green plants and spray only the plants. A machine­vision guided precision band sprayer for small-plant foliar spraying [4] demonstrated a target deposition efficiency of 2.6 to 3.6 times that of a conventional sprayer, and the non­target deposition was reduced by 72% to 99%.

Certain accurate methods for weed detection have been developed, which included wavelet transformation to discriminate between crop and weed in perspective agronomic images [5] and spectral reflectance of plants with artificial neural networks [15]. Other researchers have investigated texture features [6] or biological morphology such as leaf shape recognition [5]. So in real time for the identification and classification of crop rows in images, a lot of fast methods have been implemented [7]; some of them are based on Hough transform [8], Fourier transform [11], Kalman filtering [9] and linear regression [10]. Consequently, there are various vision systems available on autonomous weed control robots for mechanical weed removal.

Since the formulation of Perona and Malik [14] anisotropic diffusion [13] for early work on this topic), a significant amount of research has been devoted to the theoretical and practical approaches of anisotropic diffusion and related methods for image classification. Pemoa and Malik equation provides a potential algorithm for image classification and segmentation, reducing image noise without removing significant details of the image.

2010 International Coriference on Educational and Network Technology (ICENT 2010)

II. METHODOLOGY

A. Software Development The concept of the automated sprayer system (shown in

figure 5) includes camera, Central Processing Unit (CPU) and two dc pumps for spraying. The images were taken at a distance of 4 meter and at angle of 45 degree with the ground. Agriculture fields are selected for this type of study. The software is developed in Matlab. A Graphical User Interface (GUI) is developed that shows the Original image, processed image and the result of the proposed algorithm. The image resolution was 240 pixel rows by 320 pixel columns.

B. Image Pre-Processing The first stage of feature extraction algorithms is

preprocessing operation to segment all weeds from the background.

To distinguish weeds from background objects in a color image, a color segmentation image-processing step is conducted where objects are classified into one of two classes (weed and background) by their color difference in red, green and blue color space. Reference [20] indicated that weeds in field images must be carefully segmented; otherwise the feature extraction will yield unreliable results from analyzing soil and weeds. Thus, adequate image segmentation quality is necessary. One simple technique for separating pixels into weed or background class is to calculate an offset excessive green (OEG) value from the RGB image. Each pixel in the RGB image is replaced with the following calculated value:

OEG = 128 + (G - 8) + (G - R) 0)

Where R, G, B are red, green, and blue intensities of a pixel respectively. After the OEG image is generated, a threshold value is selected, based on the color differences, to separate weeds and the background.

III. ANISOTROPIC DIFFUSION WEED CLASSIFIER

Weeds are general green color, a highly irregular leaf shape and varying surface texture, and an open plant structure which contributes to its being a challenging weed to identify in the field. As the application is developed for open air natural environment, so such classifiers are of mainly interest which also enhance the image with removal of noise and preserving the significant part of image contents.

Our Weed classifier is based on anisotropic diffusion also called Perona-Malik diffusion [14]. The proposed classifier classifies the images in four categories i.e Broad Leaf, Narrow Leaf, Low Weed and Mixed.

The Anisotropic diffusion enhance the image by considering the local structures in the images to filter noise, and preserve edges, significantly increasing the signal-to­noise ratio (SNR) with no major quantitative distortions of the signal [16].

Diffusion algorithms remove noise from an image by modifying the image via a partial differential equation (PDE).

12

aI(x,y,t) = div [g�IV I I I)v I ] at

g(x)�O when x�oo (2)

Where l(x ,y,O):9\2 �9\+

is image in continuous

domain, (x, y) specifies spatial position in that image while t

is time parameter. IIV'III denotes the gradient magnitude

and g�IV III) is diffusion coefficient, controls the rate of

diffusion and is chosen as a function of the image gradient use to preserve edges.

The resultant image I(x , y, t) from anisotropic

diffusion is equal to the derivation from convolution of

original 10 (x, y) image with Gaussian kernel

G (x, y, t) of variance t [14]:

I(x,y,t)=Io (X,y)*G(X,y,t) 0)

With initial condition I( x , y, t) = 10 (x, y)

An economic chemical application threshold was considered in the simulation test derived from the summation Sj of intensities sum of resultant image.

M-IN-I

Sj = LLI(x,y, t) x=o y=O

Since information about weed numbers in unit area and average weed size (age) could be used to make the decision to skip some low weed density control zones or to decide between multiple application rates for different weed infestation levels.

IV. RESULTS AND DISCUSSION

Figure 2 and 3 shows the resultant graphs, and anisotropic diffused images. The algorithms give 100% accuracy to detect the presence or absence of weed cover. For areas where weeds are detected, results show over 97.6% classification accuracy over 200 sample images with 100 samples from each class as shown in Table 1. Average time taken by the algorithm is 0.05 m.s.

The proposed algorithm give is 95 % accuracy in classification of different leaf textures.

V. CONCLUSION

A real-time weed control system is developed and tested in the lab for selective spraying of weeds using vision recognition system. In this paper, feature extraction based system for weed classification and recognition is developed.

2010 International Conference on Educational and Network Technology (ICENT 2010)

The system shows an effective and reliable classification of images captured by a video camera.

TABLE I. RESULTS OF THE WEED CLASSIFIER

Weed Type Results Found Correct % Broad Weed 96.9 %

Narrow Weed 98.4 % Little or No Weed 100%

VI. FUTURE WORK

In this paper weed image, which has one dominant weed species can be classified reasonably accurate. But the case of more than one weed classes cannot be accurately classified. Further research is needed to classified mixed weeds. One way is to break the image into smaller region. With smaller regions, there will be less possibility to fmd more than one weed classes in this small region.

VII. REFRENCES

[1] Chancellor, W. J. & Goronea, M. A 1993. Effects of spatial variability of nitrogen, moisture, and weeds on the advantages of site­specific applications for wheat. Transactions of the American Society of Association Executives 37(3): 717-724.

[2] Shearer, S. A and P. T. Jones. 1991. Selective application of post­emergence herbicides using photoelectrics. Transactions of the Transactions of American Society of Association Executives

34(4): 1661-1666.

[3] J. E.Hanks 1996. Smart sprayer selects weeds for elimination. Agricultural Research. 44(4): 15.

[4] Giles, D. K. and D. C. Slaughter. 1997. Precision band sprayer with machine-vision guidance and adjustable yaw nozzles. Transactions of the American Society of Association Executives 40(1):29-36.

[5] Manh, AG., G. Rabatel, L. Assemat and MJ. Aldon, 2001. Weed leaf image segmentation by deformable templates. J. Agric. Eng. Res.,80: 139-146

[6] Meyer, G., T. Metha, M. Kocher, D. Mortensen and A Samal, 1998. Textural imaging and discriminate analysis for distinguishing weeds for spot spraying. Trans. ASAE, 41: 1189-1197

[7] Moshou, D., E. Vrindts, D.B. Ketelaere, DJ. Baerdemaeker and H. Ramon, 2001. A neural network based plant classifier. Comput. Electron. Agric., 31: 5-16.

[8] Leemans, V. and M.F. Destain, 2006. Application of the Hough Transform for seed row location using machine vision. Biosyst. Eng., 94: 325- 336

[9] Hague, T. and N.D. Tillet, 2001. A band pass filter-based approach to crop row location and tracking. Mechatronics, 11: 1-12

[10] Sogaard, H.T. and HJ. Olsen, 2003. Determination of crop rows by image analysis without segmentation. Comput. Electron. Agric., 38: 141-158

[11] Vioix, J.B., J.P. Douzals, F. Truchetet, L. Assemat and J.P. Guillemin, 2002. Spatial and spectral method for weeds detection and localization. EURASIP JASP, 7: 679-685

[12] Michael J. Black, Guillermo Sapiro. , David H. Marimont and David Heeger. 1998. Robust Anisotropic Diffusion. IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 7, NO. 3.

[13] D. Gabor, "Information theory in electron microscopy," Lab. Investigation, vol. 14, pp. 801-807, 1965.

[14] P. Perona and J. Malik, "Scale-space and edge detection using anisotropic diffusion," IEEE Trans. Pattern Anal. Machine Intell., vol. 12, no. 7 pp. 629-639, July 1990.

[15] Fontaine, V. and T.G. Crowe, 2006. Development of line-detection algorithms for local positioning in densely seeded crops. Canadian Biosyst. Eng., 48: 19-29

13

[16] S. Tabik, E.M. Garz'on, I. Garc'la, JJ. Fern'andez . 2006 Implementation of Anisotropic Nonlinear Diffusion for Filtering 3D Images in Structural Biology on SMP Clusters. Proceedings of the International Conference ParCo 2005, NIC Series, Vol. 33, ISBN 3-00-017352-8, pp. 727-734

[17] Sunil, K.M., P.R. Weckler and R.K. Taylor, 2007. Effective Spatial Resolution for Weed Detection. 2007 ASABE Annual International Meeting Sponsored by ASABE Minneapolis Convention Center Minneapolis, Minnesota 17 - 20 June, 2007.

[18] Wane N, Zhang. E.F, Sun Y. and D.E. Peterson, 2001, "Design of an Optical weed ditreibution for improved post emergence control decision", Weed Science, 40,546-553.

[19] Siddiqi, M.H., S.B.T. Sulaiman, I. Faye and I. Ahmad, A Real Time Specific Weed Discrimination System Using Multi-Level Wavelet Decomposition, Int. J. Agric. BioI., Vol. 11, No. 5, 2009

[20] D. M. Woebbecke, G. E. Meyer, K. Von Bargen and D. A Mortensen, "Shape features for identifying weeds using image analysis," Transactions of the ASAE, vol. 38, no.1, pp. 271-281, 1995

2010 International Coriference on Educational and Network Technology (ICENT 2010)

Decision Box

Broad Image

Anisotropic Image

Sprayer

Automatic

Sprayer Control System

Figure l. The concept of a Real-Time Specific Weed Sprayer System

Grayscale Image Narrow Image

00

Anisotropic Graph Anisotropic Image

Figure 2. Results of Anisotropic Weed Classifiers

14

Grayscale Image

Anisotropic Graph

• 0

2010 International Conference on Educational and Network Technology (ICENT 2010)

o ,.

00

Figure 3. Resultant Graphs of Anisotropic Diffusion of Various Leaf Textures

15

0 0