RECOGNITION OF PARTIALLY OCCLUDED PLANT LEAVES USING A MODIFIED WATERSHED ALGORITHM

Transcription

1 RECOGNITION OF PARTIALLY OCCLUDED PLANT LEAVES USING A MODIFIED WATERSHED ALGORITHM W. S. Lee, D. C. Slaughter ABSTRACT. Occlusion is an obstacle to two dimensional machine vision recognition of plants in natural outdoor scenes. Five modifications to the Watershed algorithm were investigated for separating occluded plant leaves, in an attempt to reduce the excessive object fragmentation associated with the original Watershed algorithm. The best modified algorithm improved the recognition of occluded tomato cotyledons and tomato true leaves improved by a factor of 3 and 2, respectively, after it was applied to occluded plant leaves in natural outdoor scenes. Two of the modified Watershed algorithms required about 11% less computation time than the original. Keywords. Object segmentation, Occlusion, Plant recognition, Watershed. One of the difficult tasks for two dimensional machine vision applications is to recognize partially occluded objects, particularly in natural scenes. Human vision can distinguish overlapped objects quite easily since it works in three dimensions. For machine vision systems to recognize overlapped objects using only a single top view, the occluded objects need to be cut apart so that they can be analyzed separately. In an early study, Lester et al. (1978) applied two boundary finding algorithms (heuristic searching and the least maximum cost technique) to white blood cell images using a graph searching method. They reported that the heuristic searching method was an effective and efficient procedure. Whittaker et al. (1987) used a modified circular Hough transform to locate partially hidden tomatoes based on shape. However, this algorithm was computationally intensive and could not be performed in real time. Shatadal et al. (1995) developed an algorithm to disconnect touching kernels using a mathematical, morphology based approach; however, they reported that the algorithm failed when the connected kernels formed a relatively long isthmus or bridge between them. One of the widely used methods for separation of occluded objects is the Watershed algorithm. Since it was first introduced as a morphological aid by Digabel and Lantuéjoul (1978), many variations of this technique have been developed. Vincent and Soille (1991) developed an efficient and fast algorithm for computing watersheds in digital grayscale images based on immersion simulation. The basic idea for this algorithm is that a digital image is considered as a Article was submitted for review in June 2003; approved for publication by the Information & Electrical Technologies Division of ASAE in April The authors are Won Suk Lee, ASAE Member Engineer, Assistant Professor, Department of Agricultural and Biological Engineering, University of Florida, Gainesville, Florida; and David C. Slaughter, ASAE Member Engineer, Professor, Department of Biological and Agricultural Engineering, University of California, Davis, California. Corresponding author: Won Suk Lee, Department of Agricultural and Biological Engineering, University of Florida, Frazier Rogers Hall, Museum Road, Gainesville, FL , phone: , ext. 227; fax: ; e mail: wslee@ufl.edu. topographical surface, and the value of each pixel represents the elevation at that point (fig. 1). Object boundaries are defined as higher elevations and regions within objects as lower elevations. Each local elevation minimum is viewed as a separate catchment basin. An imaginary hole is pierced at each local minimum (fig. 1e) in the topographical surface, and then the entire surface is flooded from below with water through the imaginary hole, beginning with the lowest local minimum. A dam is built at the point (or points) where the water from two separate catchment basins would merge, to keep the water from flooding from one basin to another. This dam is the watershed boundary (separation line) between the overlapped objects. The Watershed algorithm can be applied to either grayscale images or binary images. Russ (1990) described conditions for successful application of the Watershed algorithm: The method makes the implicit assumption that both features are actually convex, so that they should be segmented, and also assumes that the degree of touching or overlap is sufficiently small that there is a valley between the peaks at the center of each feature in the brightness coded distance map. Casasent et al. (1996) applied the binary Watershed algorithm to separate pistachio nuts in X ray images and reported that 250 of the 253 clusters were segmented correctly. However, it was not implemented in real time. They also reported over segmentation (excessive object fragmentation) problems if the boundary was irregular or complex. In order to make watershed lines thin, while straightening them to produce identical watershed lines regardless of object orientation, Orbert et al. (1993) proposed a modified Watershed algorithm that fused the catchment basins. They provided two example applications and reported that the new algorithm worked well. Lee et al. (1999) found that a robotic vision system for weed detection based on the morphological features of elongation and compactness could correctly identify, in real time (0.34 seconds per image), 73.1% of tomatoes and 68.8% of weeds from a validation set of field images taken in commercial tomato fields. One of the problems encountered by Lee et al. (1999) was an inability to correctly recognize plant leaves due to occlusion. Transactions of the ASAE Vol. 47(4): American Society of Agricultural Engineers ISSN

2 (a) (b) (c) (d) (e) (f) Figure 1. Watershed method for separating occluded objects: (a) original 3D scene, (b) top view of (a), (c) 2D binary image of (b), (d) distance function of image (c), (e) Watershed catchment basins of (d), and (f) separated binary objects. Figure 2a shows an example image of a tomato seedline with partially occluded tomato leaves. The majority of these leaves could not be recognized by the algorithm of Lee et al. (1999) due to occlusion (fig. 2b). Occluded objects need to be separated before their morphological features can be determined using 2 D vision because occluded objects appear as a single object in a binary image (fig. 2c), pro ducing an unusual set of feature values. Figures 2d through 2f show the application of the Watershed algorithm, developed by Vincent and Soille (1991), to the occluded leaves. The application of the Watershed method starts with the distance function (eq. 1), which is defined as a function that associates a value, inversely proportional to its distance to the background, for every pixel (fig. 2d): dist A (x, y) = 255 min{l(p)} (1) where x, y = pixels in image A P = paths between x and y (P should be completely included in image A) l(p) = length of P. Thus, the farther an object pixel is from the object boundary, the lighter ( lower ) the pixel is in figure 2d. Figure 2e shows the level lines (similar to topolines) of the distance function. Figure 2f shows Watershed lines generated to prevent catchment basins (local minima shown in fig. 2d) from merging in the flooding step. Figure 2g shows correctly recognized tomato leaves in gray due to reapplication of the algorithm by Lee et al. (1999) after separation by the Watershed algorithm. In this figure, some of the true leaves are still incorrectly recognized as weeds (shown in black). However, four additional tomato leaves unrecognized in figure 2b are now recognized. The Watershed method has two drawbacks: over separation of objects, and being too computationally intensive.- Over separation is excessive object fragmentation, where a single object is incorrectly cut into separate pieces. Beucher and Meyer (1992) observed that the over segmentation produced by direct construction of the watershed line is due to the fact that every regional minimum becomes the center of a catchment basin. More specifically, regional minima 1270 TRANSACTIONS OF THE ASAE

3 (a) (b) (c) (d) (e) (f) (g) Figure 2. Watershed algorithm application example: (a) grayscale version of a true color image of the seedline in a commercial processing tomato field, (b) classification (correct = gray, incorrect = black) of leaves in image (a) without application of the Watershed algorithm, (c) color segmented binary image of (a), (d) distance function of image (c), (e) level lines of image (d), (f) Watershed lines in black, and (g) improved classification of leaves in image (a) after application of the Watershed algorithm. (a) (b) Figure 3. Example of excessive object fragmentation by the Watershed algorithm due to disconnected local minima: (a) original image, (b) example of excessive fragmentation, and (c) distance function applied within the small, rectangular region A indicated in (b). (c) Vol. 47(4):

4 sometimes occur as two or three connected components in a single object, producing over separation when the Watershed algorithm is applied. For example, figure 3 shows an example of excessive object fragmentation due to discon nected local minima. When the distance between the local minimum and the background is calculated, it is influenced by the structure of the object. The leaf in the lower right corner of this image, for example, has six pixels in the fifth row (fig. 3c), none of which has a value of 253, creating two local minima in the leaf with values of 253. If there were a pixel in the dotted square a position ( ), then the pixel in dotted square 254 ( ) would become 253, or if there were a pixel in the solid square b position ( ), then the pixel in solid square 254 ( ) would become 253. In either of these two cases, the object would have one connected local minimum and would not be excessively fragmented. Thus, boundary uncertainty can have a significant impact on excessive object fragmentation by the Watershed algorithm. OBJECTIVE The objective of this research is to explore the feasibility of identifying partially occluded crop plants and weeds in natural outdoor scenes using the Watershed algorithm with different modifications. MATERIALS AND METHODS True color images of juvenile processing tomato plants from California commercial tomato fields were used to evaluate modified Watershed algorithms. The images were acquired using the robotic weed control system described by Lee et al. (1999) and stored as 24 bit RGB true color images, pixels in size, representing a cm region of seedline. They generally contained tomato and weed plants at the cotyledon and first true leaf stages. The true color images were converted to binary images, as described by Lee et al. (1999), before the Watershed algorithm was applied to separate partially occluded plant leaves. The basic Watershed algorithm, referred to as the original Watershed algorithm in this research, was the one developed by Vincent and Soille (1991). The following five modifications to the original Watershed algorithm were compared with the original. All computations were conducted with a Dell Dimension XPS Pro200n computer equipped with a 200 MHz Pentium Pro CPU. The following modifications of the Watershed algorithm are proposed in this research to explore the feasibility of identifying partially occluded crop plants and weeds in natural outdoor scenes. W0 is the original Watershed algorithm, and W1 to W5 are the modified ones: W0 Original with no modifications W1 Morphological opening W2 Pre flooding W3 Morphology criterion with pre flooding W4 Morphology criterion with opening W5 Boundary concavity criterion with opening. MODIFICATION W1: MORPHOLOGICAL OPENING A classical solution to avoid excessive object fragmentation is to slightly smooth the distance image by performing a morphological opening operation after the distance image is made but before the flooding step. Opening is defined as erosion followed by dilation and is basically a smoothing operation (Jain, 1989). Applying the opening operation to local minima increases the chance for objects that are prone to excessive fragmentation to contain only a single minimum, thus reducing excessive fragmentation. In order to obtain a larger connected local minimum, the opening operation was applied to those pixels in the distance image with values up to a certain height (e.g., h min + k, where h min is a minimum height or value in the distance image for an object where k is found by trial and error). The height (h) is defined as pixel value in the distance image. The area resulting from the opening operation was used as a new local minimum. After opening, the pixel locations of the opened area were given the pixel value of h min + k. Then, the distance function was resorted by increasing gray level, and the cumulative frequency distribution of the distance function was calculated again. The original Watershed algorithm was then applied. This modification is illustrated using the image of a tomato cotyledon shown in figure 4a. Figure 4b shows the distance image of the cotyledon in figure 4a, where the symbols of 1, 2, 3, 4, and 5 were used for heights 251, 252, 253, 254, and 255, respectively. Figure 4c shows two disconnected local minima (h min = 251), which would cause excessive object fragmentation in the original Watershed algorithm. If these minima were opened once (fig. 4d), both were removed, failing to create a single connected local minimum. Figure 4e shows the local minima as defined by pixels with values of 251 and 252, and figure 4f shows the area resulting from application of the opening operation to the local minima shown in figure 4e. The resulting area shown in figure 4f still contains two disconnected (4 connectedness) minima. Figure 4g shows the local minima as defined by pixels with values of 251, 252, and 253 in figure 4b. Figure 4h shows the area resulting from application of the opening operation to the local minima shown in figure 4g, and is now a smoothed single connected local minimum. For this object, the area resulting from the application of the opening operation to the new local minima, as defined by pixels with values up to 253 (h min + 2), has prevented the object from being excessively fragmented. Since long objects (e.g., tomato cotyledons) tend to be excessively fragmented, the length of the major axis (MJX, in pixels) of the object was used as a criterion to determine at what height the opening operation was performed prior to application of the Watershed algorithm. The MJX/opening criterion (determined by trial and error from an independent set of representative images) was: Condition Operation 0 < MJX 20 No opening 20 < MJX 35 Opening using h min < MJX 80 Opening using h min < MJX Opening using h min + 3 The area resulting from the opening operation was used as the new local minimum. MODIFICATION W2: PRE FLOODING The second modification was to use a pre flooding operation to combine the local minima of an object before the Watershed algorithm was applied. Pre flooding was defined as raising the local minima pixel values in an object to a pre determined height (pixel value in the distance image), if the pixel values were less than the pre determined 1272 TRANSACTIONS OF THE ASAE

5 (a) (c) h = 251 (e) h < 252 (g) h < 253 (d) (f) (h) (b) Figure 4. Example of morphological opening of local minima: (a) grayscale version of a true color image of a tomato seedling; (b) distance image of the object in (a), where 1, 2, 3, 4, and 5 indicate 251, 252, 253, 254, and 255, respectively; (c) local minima before opening if h = 251; (d) objects in (c) after opening; (e) local minima before opening if h < 252; (f) objects in (e) after opening; (g) local minima before opening if h < 253; and (h) objects in (g) after opening. height. If an object had pixels with values less than the pre determined height, the objects were pre flooded, and then the Watershed algorithm was applied. For example, if there were two disconnected local minima in an object with a height of 250, which could potentially lead to excessive fragmentation, then the object would be flooded with a height of 251 to connect the disconnected local minima so that excessive fragmentation can be avoided. Three pre flooding heights were evaluated: 251, 252, and 253. MODIFICATION W3: MORPHOLOGY CRITERION WITH PRE FLOODING Since the appropriate level of pre flooding varied with object shape, and object shape varied greatly, it was difficult to develop a general technique for obtaining the optimum pre flooding level for each object without extensive study. Thus, instead of determining the pre flooding level for each object, it would be more practical to determine when the Watershed algorithm needs to be applied. The features such as area (AREA = the number of pixels in an object), elongation (ELG) (eq. 2) and compactness (CMP) (eq. 3) were used to explore the feasibility of determining when the Watershed algorithm needed to be applied: MJX MNX ELG = (2) MJX + MNX 16 AREA CMP = (3) 2 PERIMETER where MNX is the length of the minor axis in pixels. The AREA was useful in determining whether the Watershed algorithm should be applied to an object, since objects with a small area were generally indicative of a single object that was probably not occluded. ELG was useful in identifying long and thin objects, such as tomato cotyledons, which tended to be excessively fragmented by the Watershed algorithm. CMP can also be used to distinguish compact objects, which had a tendency toward improper separation due to their lower local minima in the distance image. Vol. 47(4):

6 Table 1. Area (pixel 2 ), elongation (ELG), compactness (CMP), and number of concave regions (NCR) for occluded and non occluded plant objects. Non Occluded Group [a] Occluded Group [b] Area ELG CMP NCR Area ELG CMP NCR Mean SD [a] The number of objects in the non occluded group was 192. [b] The number of objects in the occluded group was 55. In order to determine the proper range of feature values (AREA, ELG, and CMP) for occluded and non occluded leaves, an independent set of 36 training images was chosen randomly from field images, and plant leaves in this training set were divided into two groups: non occluded and occluded. Table 1 shows the mean (µ) and standard deviation (σ) values for AREA, ELG, and CMP of these two groups in the training set. Then feature thresholds were calculated using Bayes rule (Duda and Hart, 1973) for AREA, ELG, and CMP in order to classify non occluded and occluded objects. The ELG distributions of the two groups (non occluded and occluded) were greatly overlapped. Therefore, ELG was omitted from the method since it was not useful in determining when the Watershed algorithm should be applied. Defining T AREA as an AREA threshold separating the non occluded objects from the occluded objects, T AREA was determined using the values in table 1 and the following equation: TAREA µ NON OCCLUDED µ OCCLUDED TAREA = (4) σnon OCCLUDED σoccluded The thresholds calculated for AREA and CMP using equation 4 were and 0.563, respectively. The following criterion was used to determine when the Watershed algorithm should be applied: Criterion 1: If AREA > and CMP < 0.563, then apply the Watershed algorithm to the object. Otherwise, do not apply the Watershed algorithm. MODIFICATION W4: MORPHOLOGY CRITERION WITH OPENING In modification W4, the morphology criterion shown in Criterion 1 was used to determine when to selectively apply the morphological opening operation to the distance image. When an object satisfied equation 1, modification W1 was selectively applied to that object. MODIFICATION W5: BOUNDARY CONCAVITY CRITERION WITH OPENING The number of concave regions (NCR) along the boundary of an object was used as a criterion for predicting object occlusion since partially occluded objects tended to have more concave regions of a certain size along their boundary than non occluded objects. The boundary curvature of each object was calculated prior to Watershed application, and the NCR was determined by counting the number of boundary regions containing negative curvature values. The Watershed algorithm was only applied when the NCR of an object met or exceeded a threshold value (T CCAVE ). The curvature of each object was calculated at boundary pixel locations using the method of Lee (1998), which determines the rate of change of the polar angle of the unit vector tangent to the object boundary using finite differences. A gap size of 5 pixels and a segment size of 5 pixels were used in the finite difference calculation to optimize the method for this application (Lee, 1998). The gap was defined as the number of boundary pixels between two consecutive segments and was used to tune the discrete derivative to accentuate concave regions of a desired size. The segment was defined as a group of contiguous pixels used to estimate the average coordinates of a boundary point and was used to reduce the adverse effects of boundary uncertainty on curvature calculations. Table 1 shows the mean and standard deviation of NCR for non occluded and occluded groups using the same 36 sample images used previously in modification W3. Based on the curvature distributions of the two groups, a T CCAVE of 4 (determined in the same fashion as T AREA ) was found to be an appropriate cutoff to determine when the Watershed algorithm should be applied. The curvature criterion was: Criterion 2: If the number of concavities > 4, then apply the W1 algorithm to the object. Otherwise, do not apply the Watershed algorithm. ASSESSMENT OF WATERSHED METHODS The ability of the original and modified Watershed methods to accurately separate partially occluded tomato and weed leaves were assessed using a set of five validation images (fig. 5a). These validation images were selected because they represent the range of occlusion conditions observed in California processing tomato fields during the first few weeks after seedling emergence, and were independent from any of the training images used to develop the Watershed modifications. To quantify the performance of each Watershed modification, the numbers of leaves excessively cut (m1), occluded leaves not separated (m2), and occluded leaves properly separated (m3) were determined after application of each method to the five images shown in figure 5b. For the original Watershed algorithm (W0), m1 = 15, m2 = 2, and m3 = 4 for the five images in figure 5c. For W0, the total number of leaves affected by the algorithm (m4 = m1 + m2 + m3) was 21. The performance of each algorithm was defined as: m3 Separation Performance = (5) m4 Since the objective of this research was to distinguish partially occluded plant leaves using a real time weed control system, the execution times for the original and the modified algorithms were measured using a 200 MHz Pentium Pro processor. The execution time for each Watershed algorithm was measured from the step of generating the distance function of the occluded binary leaf to the step of obtaining the separated leaves. In this timing task, the second image (II) from the top in figure 5b was used as the test image since the number of leaves in that image was typical in the study. To assess the performance on plant classification rate, the best modified Watershed algorithm was applied to 12 field images (different from the previous 5 test images) selected at random. When occluded objects, excessively separated by 1274 TRANSACTIONS OF THE ASAE

7 I II III IV V (a) (b) Figure 5. Examples of excessive cutting (solid arrows), inadequate separation (dotted arrows), and proper separation (star) by the original Watershed algorithm: (a) raw images, (b) binary images, and (c) results by the original Watershed algorithm. (c) Vol. 47(4):

8 the original algorithm, were properly separated by a modified Watershed algorithm, the plant leaves were labeled with numbers in the figures. For each image, the plant identification criterion shown in Criterion 3 (Lee et al., 1999) was applied to each image with and without Watershed pretreatment, and the numbers of correctly and incorrectly recognized objects were counted: Criterion 3: If < ELG < and < CMP < 1.073, then identify as tomato leaf. Otherwise, identify as weed leaf. RESULTS AND DISCUSSION The original Watershed algorithm tended to produce excessive object fragmentation when applied to the occluded tomato seedling scenes (fig. 5c). This was due to the fact that every regional minimum became the center of a catchment basin, and some object shapes resulted in multiple regional minima. The original algorithm produced 15 excessively fragmented leaves, 2 not separated leaves, and 4 properly separated leaves, giving a separation performance of 19%. All Watershed modifications did a better job of correctly separating occluded leaves than the original Watershed algorithm, with separation performances ranging from 38% to 57% (table 2). The modifications can be divided into two groups: those improving the Watershed algorithm (W1 and W2), and those applying the Watershed algorithm selectively (W3, W4, and W5). Many instances of excessive cutting were avoided with modification W1 (labeled 1 to 8 in fig. 6) since it connected local minima into larger connected components. However, some objects were still excessively fragmented, and the opening operation could not remove all instances of excessive fragmentation. The appropriate level of pre flooding used in modification W2 varied with object shape. The typical local minima in the distance image for tomato cotyledons was about 250 due to their long and thin shape. Other leaves, including true leaves and occluded leaves, had local minima as low as about 230, depending on their size. If the pre flooding level for modification W2 was set too low, the operation did not reduce over cutting, and if set too high, the objects were not separated properly. For example, the pre flooding level of 251 produced 11 over separated objects, 3 uncut leaves, and 7 properly separated leaves (fig. 7a), whereas the pre flooding level of 253 produced no excessively separated leaves, and 12 properly separated leaves with 9 uncut leaves (fig. 7c). Method Table 2. Separation performance of original and modified Watershed algorithms. No. of Over Separated Leaves No. of Uncut Occluded Leaves No. of Properly Separated Occluded Leaves Separation Performance (%) Execution Time (sec) W W W W W W W W Spatial resolution and object boundary uncertainty affected the performance of the W2 technique when applied to thin objects such as tomato cotyledons. At a spatial resolution of 0.45 mm/pixel, the presence or absence of a single pixel along the boundary of a cotyledon (fig. 3c) could lead to improper cutting by the Watershed algorithm, because, at low resolution, a single pixel could disconnect the local minima into more than two disconnected regions. The convex object assumption of the Watershed algorithm was violated when boundary uncertainty caused local concavities. The gravity of this violation was greatest for thin objects such as tomato cotyledons. There may exist a minimum spatial resolution for improved pre flooding performance in this case; however, more study is needed to investigate the relationship among the pre flooding technique, spatial resolution, and boundary uncertainty. The W3, W4, and W5 modifications selectively applied the Watershed method to objects. The morphology criteria outperformed the boundary curvature criterion; however, none of them had improved separation performance over W1 and W2 (table 2). While some leaves benefited by the criteria and were not excessively separated, some occluded leaves were skipped by the criteria and were not properly separated. When two leaves, with morphologies that tend to be excessively separated, are occluded, criterion selection tends to fail, because, in this case, errors always occur whether the modified Watershed algorithm is used or not. The main advantage of selective application was the reduction in computation time. Modifications W3 and W4 had similar separation performances to W2 and W1, respectively, but required about 15% less time to complete. However, only W3 and W5 required less computation time (11%) than the original Watershed algorithm, and none of the modified Watershed methods could be implemented in real time using a 200 MHz Pentium Pro CPU. Despite its time limitation, the Watershed algorithm improves the overall identification rate for occluded objects. Figure 8 shows correctly identified tomato and weed leaves (in gray) in the five images from figure 5a before and after modification W3 was applied. In these five images, 15 tomato leaves were correctly recognized after the Watershed algorithm application, which were incorrectly recognized as 7 occluded leaves without the Watershed algorithm. When modifications W1 and W3 were applied to the additional twelve field images, the identification rate of tomato plants increased. The plant recognition rates without any Watershed pretreatment and with modifications W1 and W3 are shown in table 3 for the twelve additional field images. Watershed pretreatments improved the recognition rates for occluded tomato cotyledons and tomato true leaves when compared to application of the Watershed method without any pretreatment. Recognition rates improved about 3 times for tomato cotyledons and 2 times for tomato true leaves. Thus, if the Watershed algorithm could be implemented in real time, it would greatly improve the recognition of occluded plant leaves when leaf morphology based feature recognition is used. Object Table 3. Plant recognition rates of tomato leaves and weeds. No Watershed Pretreatment (% recognized) W1 Pretreatment (% recognized) W3 Pretreatment (% recognized) Tomato, cotyledons Tomato, true leaves Weeds TRANSACTIONS OF THE ASAE

9 1 2 I II III IV V Figure 6. Watershed algorithm modified with the opening operation (W1). Vol. 47(4):

10 1 I 2 3 II III IV V (a) Pre flood level: 251 (b) Pre flood level: 252 (c) Pre flood level: 253 Figure 7. Watershed algorithm modified by pre flooding (W2) TRANSACTIONS OF THE ASAE

11 I II III IV V (a) Without watershed algorithm (b) With W3 watershed algorithm Figure 8. Plant recognition results without and with Watershed pretreatment (correct = gray, incorrect = black). Vol. 47(4):

12 CONCLUSION To improve the recognition of partially occluded plant leaves in a robotic weed control application, modified Watershed algorithms were implemented to cut apart occluded leaves in binary images, as a preprocessing step. Watershed modifications were investigated to evaluate their ability to reduce the tendency of the Watershed algorithm toward over cutting. The best techniques improved the separation performance by a factor of 3. Direct modifications to the Watershed algorithm, by either smoothing the distance image or adding a pre flooding step, produced results that were better or comparable to selective application of the Watershed algorithm based on leaf morphology. While none of the techniques could be implemented in real time in software using a 200 MHz Pentium Pro CPU, the execution times were about 11% less when the Watershed algorithm was selectively applied. The Watershed pretreatment had a positive effect on plant recognition rates. When leaf morphology was used as the basis for machine vision recognition, recognitions of occluded tomato cotyledons and tomato true leaves were about 3 and 2 times greater, respectively, with application of the modified Watershed pretreatment than without any Watershed pretreatment. The Watershed algorithm may not be ideal for separating occluded objects with long and thin shapes common to tomato seedlings. When the local minima were composed of two or more disconnected regions, the basic problem of over cutting was exacerbated by image noise when the leaf was only 4 or 5 pixels wide. Further study is needed to investigate the relationship between disconnected local minima, spatial resolution, and boundary noise. ACKNOWLEDGEMENT This research was supported by the Florida Agricultural Experiment Station and a grant from the California Tomato Research Institute, and approved for publication as Journal Series No. R REFERENCES Beucher, S., and F. Meyer Chapter 12: The morphological approach to segmentation: The watershed transformation. In Mathematical Morphology in Image Processing, E. R. Dougherty, ed. New York, N.Y.: Marcel Dekker. Casasent, D., A. Talukder, W. Cox, H. T. Chang, and D. Weber Detection and segmentation of multiple touching product inspection items. In Proc. Optics in Agriculture, Forestry, and Biological Processing II, Vol. 2907, G. E. Meyer, and J. A. DeShazer, eds. Bellingham, Wash.: SPIE. Digabel, H., and C. Lantuéjoul Iterative algorithms. In Proc. 2nd European Symp. Quantitative Analysis of Microstructures in Material Science, Biology, and Medicine, Caen, France, Oct J. L. Chermant, ed. Stuttgart, Germany: Riederer Verlag. Duda, R. O., and P. E. Hart Pattern Classification and Scene Analysis. New York, N.Y.: John Wiley and Sons. Jain, A. K Fundamentals of Digital Image Processing. Englewood Cliffs, N.J.: Prentice Hall. Lee, W. S Robotic weed control system for tomatoes. Unpublished PhD diss. Davis, Cal.: University of California, Davis, Department of Biological and Agricultural Engineering. Lee, W. S., D. C. Slaughter, and D. K. Giles Robotic weed control system for tomatoes. Precision Agric. 1(1): Lester, J. M., H. A. Williams, B. A. Weintraub, and J. F. Brenner Two graph searching techniques for boundary finding in white blood cell images. Computers in Biology and Medicine 8(4): Orbert, C. L., E. W. Bengtsson, and B. G. Nordin Watershed segmentation of binary images using distance transformations. In Proc. Nonlinear Image Processing IV, Vol. 1902, E. R. Dougherty, J. Astola, and H. G. Longbotham, eds. Bellingham, Wash.: SPIE. Russ, J. C Computer Assisted Microscopy: The Measurement and Analysis of Images. New York, N.Y.: Plenum Press. Shatadal, P., D. S. Jayas, and N. R. Bulley Digital image analysis for software separation and classification of touching grains: I. Disconnect algorithm. Trans. ASAE 38(2): Vincent, L., and P. Soille Watersheds in digital space: An efficient algorithm based on immersion simulations. IEEE Transactions on Pattern Analysis and Machine Intelligence 13(6): Whittaker, A. D., G. E. Miles, O. R. Mitchell, and L. D. Gaultney Fruit location in a partially occluded image. Trans. ASAE 30(3): TRANSACTIONS OF THE ASAE