Automated Counting of Rice Planthoppers in Paddy Fields Based on Image Processing

Transcription

1 Journal of Integrative Agriculture Advance Online Publication 2014 Doi : /S (14) Automated Counting of Rice Planthoppers in Paddy Fields Based on Image Processing YAo Qing 1, XIAN Ding-xiang 1, YANG Bao-jun 2, Tom Dietterich 3, LIU Qing-jie 1, DIAO Guang-qiang Diao 1 and TANG Jian 2 1 School of Information Science and Technology, Zhejiang Sci-Tech University, Hangzhou , P.R.China 2 State Key Laboratory of Rice Biology, China National Rice Research Institute, Hangzhou , P.R.China 3 School of Electrical Engineering and Computer Science, Oregon State University, Corvallis, OR, 97330, USA Abstract A quantitative survey of rice planthoppers in paddy fields is important for assessing the population density and forecasting decisions. Manual rice planthopper survey methods in paddy fields are time-consuming, fatiguing and tedious. This paper describes a handheld device for easily capturing planthopper images on rice stems and an automated method for counting rice planthoppers based on image processing. The handheld device consists of a digital camera with WiFi, a smartphone and a stretchable pole. The surveyor can use the smartphone to wirelessly control the camera, which is fixed on the front of the pole, to photograph planthoppers on rice stems. For the counting of planthoppers on rice stems, we adopt three layers of detection that involve the following: (a) the first layer of detection is an AdaBoost classifier based on Haar features, (b) the second layer of detection is an SVM classifier based on HOG features, (c) the third layer of detection is the threshold judgment of the three features. We evaluated this method for counting whiteback planthoppers (Sogatella furcifera) on rice plant images and achieved an 85.2% detection rate and a 9.6% false detection rate. Our method is feasible for the assessment of the population density of rice planthoppers in paddy fields. Key words: insect counting, rice planthoppers, handheld device, AdaBoost classifier, SVM classifier, image features Introduction Rice planthoppers are one major group of rice pests. This group includes the following three important species: the brown planthopper (Nilaparvata lugens(stål)), the whiteback planthopper (Sogatella furcifera (Horváth)) and the small brown planthopper (Laodelphax striatellus (Fallén)). These species often cause serious losses in rice yield (Pathak and Khan, 1994). A quantitative survey of rice planthoppers in paddy fields is important for assessing the population density and for making forecasting decisions concerning the rice planthoppers. In most rice-growing countries, the quantitative survey of rice planthoppers currently undertaken in paddy fields usually involves one of the following three methods (China national standardization management committee, 2009). The first method is the most often used. It involves two people, one surveyor and one recorder 1

2 (Fig. 1). The surveyor first goes into a paddy field, stoops and flaps the stems of one cluster of rice several times using one hand, making the planthoppers on the rice stems drop onto an enamel plate held by the other hand. He then stands up, ignores non-planthopper entities (such as spiders, midges, mud) and quickly counts the rice planthoppers present before the planthoppers hop away from the enamel plate. The recorder on the ridge then writes down the number of planthoppers on a recording sheet. The enamel plate can be glued to avoid the escape of planthoppers. The second method is to either use a cage to cover one cluster of rice or to use a sweep net to sweep the rice. A surveyor then visually counts the number of planthoppers in the cage or net. The third method is to visually count the planthoppers on the rice stems. During the quantitative survey of rice planthoppers in paddy fields, the surveyor needs to continuously stoop down, stand up and count the rice planthoppers. These manual methods for counting rice planthoppers in paddy fields are time-consuming and tedious. If the number of rice planthoppers on the enamel plate or in the net is over fifty, surveyors usually estimate the number of the planthoppers according to personal experience, which often leads to a low accuracy rate. Fig. 1 Manual surveying of rice planthoppers in paddy fields by flapping the planthoppers onto an enamel plate for counting The traditional methods of pest identification and counting mostly depend on human visuals. These methods can bring about visual fatigue, are time consuming and can provide inaccurate data. To address these problems, computer vision technology has been used to automatically identify and count greenhouse pests and crop pests. Most of them use images from yellow sticky traps (Cho et al. 2007; Qiao et al. 2008; Bechar et al. 2010; Kumar et al. 2010; Martin et al. 2011; Zhang et al. 2006). In these studies, yellow sticky traps are used to trap pests and are photographed by a camera or scanner for automated counting based on image processing methods. It is easier to segment pest areas from sticky traps than from plant leaves and crop fields due to the yellow background that is used. However, using sticky traps is not economical for monitoring pest populations because the sticky traps often need to be replaced at regular intervals. Some studies use images from the leaves themselves (Shariff et al. 2006; Zhao et al. 2007; Boissard et al. 2008; Huddar et al. 2012; Mundada et al. 2013). However, we found the background of these images is relatively simple and it is easily to segment these pests from the leaves by pests color and texture features. The environment of rice planthoppers in paddy fields is complex and changing (Fig. 2). It is 2

3 challenging to remove the background from the planthopper by common segmentation methods without a strong generalization method (such as threshold-based segmentation and clustering). Park et al. (2003) photographed brown planthoppers in paddy fields with a camera and used some image processing methods (image decomposition, top-hat transformation, thresholding and filtering) to count N. lugens on rice plants for the estimation of planthopper density. Park set the threshold of planthopper area in the binary images to determine whether the region contained a planthopper. This method hardly removes noise of similar size to the planthoppers. In addition, Park s method can miss some planthoppers because different planthopper instars on the same cluster of rice may be of different sizes and colors. Zou et al. (2012) designed a recognition system of pests with digital signal processor for counting the rice planthoppers trapped by lamp on a white cloth. But the method couldn t replace the field survey of rice planthoppers. The goal of our research is to provide a rapid and easy system for the automated counting of rice planthoppers in paddy fields. To achieve this goal, we developed a system that combines a handheld device for photographing images of rice planthoppers on rice stems with a software system for automated counting the planthoppers in the images. To photograph planthopper images without stooping down, the handheld device was developed using a digital camera, a mobile phone and a stretchable pole. The size of rice planthoppers is small, from 1 to 5 mm, and the paddy field environment is complex and changing. Each image from the paddy fields may contain rice, rice planthoppers, other insects, water, dead leaves, dirt, weeds, disease spots and water reflection (Fig. 2). It is difficult to remove such a complex background using common image segmentation methods. We didn t segment the background from images to count the planthoppers, but proposed three layers of detection algorithm to detect and count the planthoppers in these images. Fig. 2 Rice planthoppers on rice stems and the detected area in the white box 3

4 Results Collection of the rice images by our handheld device To reduce the labor intensity and to improve efficiency, a handheld device was developed for easily collecting images of rice planthoppers on rice stems (Fig. 3). With this device, the surveyor does not need to stoop down to collect rice planthoppers onto an enamel plate by tapping the rice for visual counting. Instead, he only holds the pole with one hand and places the camera close to the rice stems. The smartphone is held with the other hand. It can remotely connect to the digital camera by WiFi and control the camera by the remote viewfinder application and shares a low-resolution version of the image in the camera lens. The surveyor previews the rice image on the smartphone screen and can move the pole until the camera finds a good view. He then uses the smartphone to control the camera to take pictures. Fig. 3 A handheld imaging device for capturing images of rice planthoppers on rice stems We captured 92 rice images in the paddy fields using our handheld device. In these images, only S. furcifera were found. The rice growth stages were between the tillering stage and the heading stage. S. furcifera instars between the 2-instar nymph and the adult stage were found. Images were divided into four levels according to the number of planthoppers found. The first level was from 0 to 10 planthoppers, which indicated a low density of planthoppers. The second level was from 11 to 20 planthoppers, which indicated a median density of planthoppers. The third level was from 21 to 30 planthoppers, which indicated a high density of planthopper. The forth level was over 30 planthoppers, which indicated a very high density of planthoppers. Table 1 summarizes the number of tested images found at each level. Table 1 Planthopper density levels in the images tested in this study Level No. of planthoppers No. of images >

5 Planthopper detection results using three layers of detection In our work, we propose three layers of detection algorithm to count the rice planthoppers on rice stem. The detailed steps of detection are described in section 5. Table 2 summarizes the detection results from the four levels of planthopper density using three layers of detection. All images were tested using the first layer of detection, an AdaBoost classifier. The detection rates (see the definition in section 5.2) were high, from 90.9% to 95.5%. The AdaBoost classifier based on Haar features had detected the presence of rice planthoppers well. However, the false detection rates (see the definition in section 5.2) were very high, from 77.6% to 97.7%. This result means that many non-planthopper forms were falsely detected using the AdaBoost classifier. To reduce the false detection rate, a second layer of detection, an SVM classifier based on HOG features, is used to identify the sub-widows detected by the AdaBoost classifier. We found that the false detection rates in all levels decreased significantly. The detection rate also decreased slightly, which means that some planthoppers are mistaken as non-planthopper forms by the second layer of detection. To further reduce the false detection rate, the third layer of detection based on the threshold judgment of three features was applied to the sub-windows as detected by the SVM classifier. The false detection rates at all levels decreased significantly. However, the detection rates also decreased slightly. In ninety-two images, we finally obtained a detection rate of 85.2% and a false detection rate of 9.6%. In the low-density planthopper images, the detection rate is the smallest and the false detection rate is the highest. This result is mainly because the total number of planthoppers on one image is small, resulting in a low detection rate and a high false detection rate. Table 2 Detection of planthoppers at the different density levels using three layers of detection Level (planthopper number in one image) Detection layer 1 (0-10) 2 (11-20) 3 (21-30) 4 (>30) Average (Detection methods) DR FDR DR FDR DR FDR DR FDR DR FDR The first layer of detection (AdaBoost based on Haar features) The second layer of detection (SVM based on HOG features) The third layer of detection (Threshold judgment of three features) DR: Detection rate (%); FDR: False detection rate (%). Discussion Pest counting is an important routine in agriculture for the estimation of pest population density and dynamics in fields that allows for precision pesticide application. At present, counting pests by human visuals is a drudgery and subjective. Due to the complex environment background of living pests, it is a big challenge to automatically identify and count them by image processing. Many researchers in pattern recognition, artificial intelligence and machine learning are developing some technologies for automating pest identification and count. It could lead to less drudgery and more 5

6 accuracy (MacLeod, et al, 2010). For reducing the labor intensity and improving the accuracy rate of planthopper field survey, we developed a handheld device and proposed an automated detection method for planthoppers in paddy fields. The detection algorithm includes three layer of detection steps: the AdaBoost classifier based on Haar features, the SVM classifier based on HOG features and the threshold judgment of the three features. This detection algorithm needs to stepwise detect the planthoppers on rice stem. We finally obtained the detection rate of 85.2% and the false detection rate of 9.6%. The result is generally satisfactory. There are a few issues that must be addressed before our method is ready for field testing. First, our handheld device only captures one side of one cluster of rice. A model should therefore be developed to predict the number of planthoppers on one cluster of rice using our counting results. Second, our classifiers are only trained using images of the whiteback planthopper Sogatella furcifera. Two other species, N. lugens and L. striatellus, often appear on rice in paddy fields together with S. furcifera and damage rice plants. In practice, the three species of planthopper should be counted respectively. So we need to train our classifiers using the three species of planthoppers in order to count them respectively. Third, the false detection rate is relatively high when the planthoppers are young (when body size is smaller). It is mainly because the young planthoppers have small image areas which could provide less image features and our classifiers couldn t identify them well. In the low density of planthoppers, we could achieve a high false detection rate when the false detection occurs. The further researches should focus on the detection of the young planthoppers and the low density of planthoppers. Finally, the rice plants would be high degree closing and the paddy fields would be complete canopy during the heading stage. The surveyor can t see the rice stem and the handheld device will touch the rice leaves, even mud. The camera on our handheld device should be equipped a waterproof cover for avoiding contamination. Additionally, in this dark situation, the quality of the images may be affected. We should add the trained examples of rice heading stage. Beyond the capturing of images of rice planthoppers in paddy fields, our handheld device has many other potential applications for the detection of pests and diseases on crops or other plants. Our detection method can be used to automate the counting of small objects in complex and variable environments when combined with other image features. Conclusion In this paper, we presented a handheld device and an automated method for counting planthoppers on rice stems in paddy fields. The handheld device can easily capture images containing rice planthoppers on rice stems. The surveyor can adjust the length of the pole and move the camera close to the rice stems using the stretchable pole. The surveyor can use the mobile phone to control the camera over WiFi to capture the planthopper images on the rice stems without continuously stooping down and standing up and visually counting the planthoppers. These images are saved on an SD card in the camera in real-time, and the automated counting of the planthoppers in the rice images is achieved using three layers of detection. Our detection methods have an 85.2% detection rate and a 9.6% false detection rate. Our methods would not only reduce labor intensity and visual fatigue in surveyors, but would also improve the accuracy of counting for rice planthoppers. 6

7 Material and Method Development of a handheld device and collection of rice images The manual survey method for counting planthoppers in paddy fields is time-consuming, tedious and subjective. To reduce the labor intensity and to improve efficiency, we developed a handheld device for easily capturing images of rice planthoppers on rice stems (Fig. 3). This device consists of the following three parts: (1) a digital camera (Samsung SH100, 14 effective Megapixels) with WiFi; (2) a smartphone (Samsung I100); and (3) a light and stretchable pole. The stretchable pole can be stretched to about 2m. The camera is fixed on the front of the stretchable pole with universal joints. The smartphone remotely connect to the camera by Wifi and control the camera to take photos by the remote viewfinder application (from Samsung Electronics CO.LTD.) installed on the smartphone. The images are saved on an SD card in the camera. Most of the planthoppers are found on rice stems. For this, we set a detection area for each image to decrease the run time and the complexity of the algorithm program. The hight and width of each image are divided into 3 x 5 equal segments, so that the image has five boxes in each row. The detected areas (Fig. 2) are the 2nd to the 4th boxes from the second row down to the bottom (third) row. It is necessary that the rice stems should be within the detected area when the rice is photographed. Counting planthoppers on rice stems using image processing methods The purpose of the software system is to automatically count the rice planthoppers on the rice stems based on image processing. We developed a detection method using three layers of detection algorithm to detect and count the planthoppers in the images. We adopted the AdaBoost classifier [Viola and Jones, 2004] as the first layer of detection. The planthoppers are detected from the complex rice background rather than removing the rice background to retain planthoppers in the images. However, we achieve a high detection rate and a high false detection rate. To reduce the false detection rate for planthoppers, the second layer of detection, which is based on HOG features and a SVM classifier, was used to further determine whether the sub-windows detected in the first step contain rice planthoppers. To remove water drops and water reflections, the third layer of detection was developed. In this step, these factors are removed using a threshold value judgment of three features after an automated removal of the background. Fig. 4 provides a flowchart for the automated counting of planthoppers on rice stems using three layers of detection. Fig.2 is used as a case for presenting the detection results from three layers of detection. The detection results are evaluated by the detection rate and the false detection rate. The detection rate is the ratio of the number of the detected planthoppers to the number of all planthoppers in an image. The false detection rate is the ratio of the number of the non-planthopper sub-windows but detected mistakenly as planthoppers to the number of all detected sub-windows. 7

8 Positive samples Negative samples Extract Haar features Training Positive samples Negative samples Extract HOG features Training Images of planthoppers on rice stems Testing AdaBoost classifier Detected sub-windows Extract HOG features Testing SVM classifier The first layer of detection The second layer of detection Removing background of planthopper sub-windows Extract three features Threshed value judgment The third layer of detection Count planthoppers Fig. 4 A flowchart for the automated counting of planthoppers on rice stems using three layers of detection based on image processing. The first layer of detection using the AdaBoost classifier based on Haar features AdaBoost is an algorithm for constructing a strong classifier from a linear combination of weak classifiers. This algorithm was proposed by Freund and Schapire (Freund and Schapire, 1997). Viola and Jones built a face detection framework using an integral image, AdaBoost and a cascade of classifiers that yielded an extremely reliable and efficient fact detector in 2004(Viola and Jones, 2004). We adopted the Viola and Jones method of face detection to detect planthoppers on rice stems. We built a collection of positive planthopper examples and negative non-planthopper examples. Eleven Haar-like features were then extracted from the positive and negative examples. These Haar features were used to train a set of weak classifiers, and an AdaBoost algorithm was used to construct a strong classifier using a linear combination of the weak classifiers. Training database The images of the rice planthoppers were captured in the paddy fields using our handheld device. From these images, we randomly cut 1800 planthopper images as positive examples and 4000 non-planthopper images as negative examples (Fig. 5; In general, the size of negative examples is not less than the size of positive examples. In our work, the sizes of the positive and negative examples are 18 by 24 pixels and 19 by 25 pixels respectively). These positive and negative examples were used as a training set for the AdaBoost classifier. 8

9 (a) Positive examples of planthoppers (b) Negative examples of non-planthoppers Fig. 5 Positive and negative examples used for training the AdaBoost classifier Extracting Haar features and training the AdaBoost classifier According to the morphologies and the locations of planthoppers on the rice stems, we selected eleven Haar-like features (Fig. 6). These features are extracted from the positive and negative examples using an integral image method to reduce the computation time [Viola and Jones, 2004] and to train the cascaded classifiers. In our work, four cascaded classifiers were combined into a strong AdaBoost classifier. The AdaBoost classifier is treated as the first layer of detection of planthoppers on rice images. (a) Edge features (b) Center features (c) Line features Fig. 6 Eleven Haar-like features Detecting planthoppers using the AdaBoost classifier The first layer of detection scans across the rice image at multiple scales. In our work, the size of the initial detection window is 18 by 24 pixels and the scale factor of the detection is 1.2. When the size of the detector is scaled up to triple the size of the initial window, the detection process stops. Multiple detections usually occur around each planthopper in a scanned image because the detector is insensitive to small changes in translation and scale. When the bounding regions overlap, we retain only the smallest bounding region. We detected planthoppers in Fig. 2 using the AdaBoost classifier (Fig. 7). Fig. 7 shows that 9

10 eighty-two sub-windows were detected as having planthoppers using the AdaBoost classifier. These sub-windows are marked by a red bounding region. We found forty-three planthoppers by visually inspecting Fig. 2. Thirty-nine of the forty-three planthoppers were detected using the AdaBoost classifier. Four planthoppers were missed. At the same time, an additional forty-three non-planthopper forms were mistakenly detected. This result indicates that we can obtain a high detection rate of 90.7% and a high false detection rate of 52.4% in this image. It will result in a low accuracy rate for planthopper counting. In the false detection sub-windows, we find that some impurities, exuviates, water, reflected light and dead leaves on the rice stems are falsely detected as planthoppers by the AdaBoost classifier. To reduce the false detection rate, we need a second detector that can reject these false detection sub-windows. The detailed method used is described in section Fig. 7 Detection of planthoppers on rice stems using the AdaBoost classifier The second layer of detection using an SVM classifier based on HOG features To reject these false detection sub-windows, we developed a second layer of detection using an SVM classifier (Cortes and Vapnik, 1995) based on Histogram of Oriented Gradient (HOG) (Dalal and Triggs, 2005) features. Extracting HOG features and training an SVM classifier Histogram of Oriented Gradient (HOG) is a feature descriptor that can count occurrences of gradient orientation in localized portions of an image. It was proposed by Navneet Dalal and Bill Triggs for human detection in 2005 (Dalal and Triggs, 2005). The HOG descriptor upholds invariance to geometric and photometric transformations, except for object orientation. It can describe the outline of a planthopper. The HOG features are extracted from planthopper sub-windows, which are detected by the first layer of detection and scaled to a consistent size of 36 by 48 pixels (Fig. 8) to train an SVM classifier. 10

11 In our work, the HOG descriptor has the following properties: grey space, R-HOG blocks, 12*16 pixel blocks of four 6 8 pixel cells, 9 orientation bins in 0-180, a block horizontal spacing stride of 6 pixels and a vertical spacing stride of 8 pixels. We achieve 900 gradient vectors from each image. The HOG features of 6000 planthopper positive examples and non-planthopper negative examples were extracted to train an SVM classifier with a radial basis kernel function. This SVM classifier is used as the second layer of detection for the rejection of false detection sub-windows. Fig. 8 Positive and negative examples and their HOG descriptors Detecting planthoppers using an SVM classifier The eighty-two sub-windows detected by the first layer of detection are also detected by the SVM classifier. Forty-five sub-windows are detected as having planthoppers (Fig. 9). Thirty-nine sub-windows with planthoppers are detected by the SVM classifier, but six out of the forty-five non-planthopper sub-windows are regarded as having planthoppers. Here, we obtain a 90.7% detection rate and a 13.3% false detection rate from the second layer of detection using the SVM classifier. (a) (b) Fig. 9 Planthopper detection using an SVM classifier based on HOG features; (a) Planthopper detection by the second layer of detection (Blue rectangles represent non-planthoppers, and red rectangles represent planthoppers); (b) Sub-windows containing planthoppers after the removal of non-planthopper sub-windows 11

12 The third layer of detection using three feature thresholds We find that some non-planthopper forms are still mistaken as planthoppers. It is because the HOG features of these forms are similar to those of planthoppers. These distracting forms are mostly water drops and water reflections. To further decrease the false detection rate, we extracted three global features of sub-windows as detected by the second layer of detection after an automated removal of the background using the Muti-Ostu method (Liao et al. 2001). We reject these distractions using the threshold values judgment of the three global features. Removing background in sub-windows We first compute two color values from each sub-window detected by the second layer of detection using the equation: Sr 2* r g Sb 2* b r where r, g, and b are red, green and blue components in RGB color place, respectively. We then compute the two threshold values T 1 and T 2 using OSTU (Liao et al. 2001). T 1 is to remove the green rice background as much as possible, and T 2 is to remove as much noise as possible while keeping the planthoppers. Finally, the background is removed using the following equation: g ( r, g, b) f ( r, g, b) S T & & S T g( xy, ) ( r, g, b) (0,0,0) other ( x, y) ( x, y) r 1 b 2 where f (x,y) (r,g,b) is the pixel value of point (x,y) in the sub-windows, and g (x,y) (r,g,b) is the pixel value of point (x,y) after the background is removed. Fig. 10 shows the results when sub-windows with planthoppers and non-planthoppers have the background removed using the Muti-Ostu method. Fig. 10. The sub-window images of planthoppers and non-planthoppers with backgrounds removed using the Muti-Ostu method Extracting the three features We selected two shape features and one color feature for the rejection of non-planthopper forms. The formulas used for the three features are as follows: 12

13 (1) D = S obj, where S obj is the area of an object and M and N are the length and width of M* N the sub-window, respectively. (2) A = W H, where W is the width of an object and H is the height of an object. (3) M = xp( x), where x is a gray value and p(x) is the probability of x. x Judgment of planthoppers using three feature thresholds We computed the three features of 1500 planthopper images and 1500 non-planthopper images from the sub-widows detected by the second layer of detection. We determined the following judgment criteria for the three features of planthoppers: 60<M<190 && 0.09<D<0.56 && 0.26<A<0.60 If the three features in one sub-window meet the above criteria, the sub-window is judged as having a planthopper. (a) (b) Fig. 11 The detection result using three layers of detection. Green rectangle: non-planthoppers; Red rectangle: planthoppers; (a) detection result using the third layer of detection; (b) sub-windows containing planthoppers The forty-five sub-windows in Fig. 9(b) that were detected using the second layer of detection are further judged using the thresholds of the three features. In total, we detected forty-one sub-windows that were regarded as having planthoppers (Fig. 11). Thirty-nine sub-windows with planthoppers were still detected and four out of six non-planthopper sub-windows were rejected by the third layer of detection. In the end, a 90.7% detection rate and a 13

14 4.9% false detection rate were achieved after the third layer of detection using the threshold judgment. Acknowledgements The authors gratefully acknowledge the support of the National Natural Science Foundation of China ( ), the National High Technology Research and Development Program of China (863 Program) (2013AA102402) and Major Scientific and Technological Special of Zhejiang province (2010C12026). Dietterich was supported by the US National Science Foundation under grant number References Bechar I, Moisan S On-line Counting of Pests in a Greenhouse using Computer Vision. In VAIB Visual Observation and Analysis of Animal and Insect Behavior. Boissard P, Martin V, Moisan S A cognitive vision approach to early pest detection in greenhouse crops. Computer and Electronics in Agriculture, 13, China national standardization management committee The standard on the forecast and survey of rice planthoppers (GB/T ). Cho J, Choi J, Qiao M, Ji C W, Kin H Y, Uhm K B, Chon T S Automatic identification of whiteflies, aphids and thrips in greenhouse based on image analysis. International Journal of Mathematics and Computers in Simulation, 1, Cortes C, Vapnik V Support-vector networks. Machine Learning, 3, Dalal N, Triggs B Histograms of oriented gradients for human detection. Computer Vision and Pattern Recognition, 1, Freund Y, Schapire R E A Decision-theoretic generalization of on-line learning and an application to boosting. Journal of Computer and System Sciences, 1, Huddar S, Gowri S, Keerthana K, Vasanthi S, Rupanagudi S Novel algorithm for segmentation and automatic identification of pests on plants using image processing. In Intl. Conf. Comp. Comm. & Networking Technologies, 1 5. Kumar R, Martin V, Moisan S Robust Insect Classification Applied to Real Time Greenhouse Infestation Monitoring. Proceedings of the 2oth International Conference on Pattern Recognition on Visual Observation and Analysis of Animal and Insect Behavior Workshop, Istanbul, Turkey, 1-4. Liao P S, Chen T S, Chung P C A fast algorithm for multilevel thresholding. Journal of Information Science and Engineering, 5, MacLeod N, Benfield M, Culverhouse P Time to automate identification. Nature, 9, Martin V, Moisan S, Paris B, Nicolas O Towards a video camera network for early pest detection in greenhouses. In Proceedings of International Conference on Endure Diversifying Crop Protection, La Grande Motte, France, Mundada1R G, Gohokar D V V Detection and classification of pests in greenhouse using image processing,. 14

15 IOSR Journal of Electronics and Communication Engineering, 6, Park Y S, Han M W, Kim H Y, Uhm K B, Park C G, Lee J M, Chon T S Density estimation of rice planthoppers using digital image processing algorithm. Korean Journal of Applied Entomology, 1, Pathak M D, Zeyaur R K Insect Pests of Rice. International Rice Research Institute, Manila, Philippines. Qiao M, Lim J, Ji C W, Chung B K, Kim H Y, Uhm K B, Myung C S, Cho J, Chon T S Density estimation of Bemisia tabaci (Hemiptera: Aleyrodidae) in a greenhouse using sticky traps in conjunction with an image processing system. Journal of Asia-Pacific Entomology, 11, Shariff A R M, Aik Y Y, Hong W T, Mansor S, Mispan R Automated identification and counting of pests in the paddy field using image analysis. Computer in Agriculture and Natural Resource, 4 th word Congress Conference, Viola P, Jones M J Robust real-time face detection. International Journal of Computer Vision, 2, Zhang J W, Wang Y M, Shen Z R Novel method for estimating cereal aphid population based on computer vision technology. Transactions of the CSAE, 9, (in Chinese with English abstract ) Zhao J, Cheng X P Field pest identification by improved texture segmentation scheme. New Zealand Journal of Agricultural Research, 5, Zou X G, Ding W M Design of processing system for agricultural pests with digital signal processor. Journal of Information & Computational Science, 15,