In the plantation sectors, weeding strategy plays a significant role in managing productivity of palm oil plantation as well as controlling palm oil quality. Currently, most plantation companies adopt manually sprayed herbicide as their weeding strategy which is known to be inefficient, labor intensive and also hazardous to the environment as well as the plantation workers. Furthermore, new regulations have been enforced to limit herbicides or pesticides concentration in drinking water leading to more environmentally responsible weeding strategies (Barnes et al., 1996). At present, herbicides are applied uniformly on the field, even though researches have shown that the spatial distribution of weed is non-uniform. If the non-uniformity of the weed spatial distribution can be identified and detected, then it would be possible to reduce herbicide quantities through efficient weeding strategy (Lindquist et al., 1998; Manh et al., 2001).
In this research, we intend to study and attempt the use of SIFT in classifying different types of weed in palm oil plantation. In controlling weed population, different herbicide formulations are used for different types of weed. In order to avoid environmental pollution due to over usage of herbicides that can contaminate the drinking water supply, it is important that the use of such formulation be controlled and the minimum amount required is being used. Consequently, an intelligent system for automated weeding strategy is greatly needed to replace manual spraying system that is able to protect the environment and at the same time, produce better and greater yields. Appropriate spraying technology and decision support systems for precision application of herbicides are available but they are unsuitable for the weed types that exist in Malaysian palm oil plantation. To date, the automatic weeding systems have demonstrated potential herbicide savings of 30 to 75% (Stafford, 1997).
Machine vision methods are based on digital images, within which, geometrical,
utilized spectral reflectance or absorbance patterns to discriminate between
narrow and broad weed. Machine vision methods have been used to show shape features
that can be used to discriminate between corn and weeds (Meyer et al.,
1994). Other studies classified the scene by means of color information (Cho
et al., 2002). Most of the reported study did not maximize and thoroughly
analyze the usage of image processing technique but rather used the general
technique on standard RGB colour and shape analysis. In the present study, the
proposed technique using SIFT, transforms the original image to the values called
key point which is found very robust to common problem in image processing such
as lighting, range of object and size of image. SIFT technique is based on scale
space fundamental theory which is known as the best technique of structural
analysis for object detection. In the image analysis study, there are two common
approaches of processing namely statistical and structural techniques. In (Victor,
2005), statistical approach has been used to analyze the weed presence in cotton
fields. It was reported that, the used of statistical approach gave very weak
detection with 15% of false detection rate. On top of that, structural approach
of SIFT technique was introduced to compare the detection and classification
rate of weed using statistical approach in (Ishak et al., 2007).
SIFT has been proven to be the most robust technique for object detection (Lowe, 2004). The SIFT key descriptor is well localized in both the spatial and frequency domains, reducing the probability of disruption by occlusion, clutter, or noise. Large numbers of features can be extracted from typical images with efficient algorithms. In addition, the features are highly distinctive, which allows a single feature to be correctly matched with high probability against a large database of features, providing a basis for object and scene recognition. Krystian and Schmid (2001) used the scale invariant technique to detect edge base features object. SIFT technique was used by Ellekilde et al. (2006) to develop rescuer robot. Due to many key points generated from an image, it is difficult to analyze the keys for feature extraction. Nowak et al. (2006) proposed Bag-of-Feature representation to make the key point easy to analyze.
In this study, since it is impractical to detect and analyze the weed according to its exact species due to large coverage area, we have addressed the problem by using a common approach, which is, by classifying the weed population into two broad classes based on leaf size. The two classes are broad and narrow weed. Classification will be based on the weed images and SIFT-C, the abbreviation for SIFT for Classification, that will be used as the main method of image analysis.
MATERIALS AND METHODS
The overall block diagram of weed classification system is shown in Fig.
1. Firstly, coloured digital images of narrow and broad weed were captured
using a C-5050 ZOOM digital camera (Olympus Digital 5.0 Mega pixel) under field
lighting condition at noon. An image resolution of 320x240 pixels was used,
which then translated into 2 pixels per mm. Images were downloaded using the
USB cable interface and stored in JPEG file format. More than 1000 images were
used in the processing and testing stage to evaluate the SIFT-C. The images
were taken under varied condition of weed such as high and low distribution.
||Overall block diagram of the weed classification method
The SIFT technique started with the implementation of the Gaussian filter by applying the scale space fundamentals theory (Tony, 1993) which involved different types of Gaussian (DoG) techniques. The pyramid level of scale space consists of a very large data and the SIFT technique has minimized it by determining its peak value. The peak value is measured by implementing the key localization technique which compared the maximum and minimum values of the neighbors from all pyramid level. Practically, the maximum and minimum values of the pyramid level can be interpreted as a distinctive edge of the object in the processed image. Next, we investigate the orientation gradient value to achieve image rotation invariant and rearrange it in the form of orientation histogram. The SIFT key descriptors produce a compact feature descriptor to describe an image using its keypoints. Generally, SIFT keypoint is mainly used for object detection. To recognize and classify an object, a matching algorithm is required. Therefore, in order to utilize the SIFT algorithm for classification purpose, we have proposed a new technique that we named as SIFT-C. The new SIFT-C algorithm is used to classify targeted object, namely the narrow and broad weed. Quelhas et al. (2005) proposed a technique for visual scene representation using quantized colour and texture local invariants features computed over regions of point of interests. The use of quantized local invariant features has been proven to provide a robust and versatile way to model images, leading to good classification (Willamowski et al., 2004), retrieval (Sivic et al., 2005) and image segmentation (Dorko and Schmid, 2003). In this study, the SIFT-C is used to derive the magnitude and orientation of the local descriptors and rearrange the histogram into a single keypoint. The feature vector set of SIFT-C is then extracted and used in the design of the classifier system.
SIFT theory: Difference of Gaussian (DoG) is the first stage of SIFT
computation. The main function is to search over all scales and images locations
to identify potential interest points that are invariant to scale and orientation
(Lowe, 2004). Firstly, the raw image with RGB colour code is converted to gray
scale and subjected to the DoG technique. The DoG is similar to the Laplace
of Gaussian technique in which the image is first smoothen by convolution with
the Gaussian kernel of certain width as shown in (1) and (2) (Lowe, 2004).
The DoG as an operator or convolution kernel is defined as (3)
The initial output of DoG produces a large pixel value at each level and each
scale space. In order to reduce the size of DoG output, the local maxima and
minima technique need to be implemented to search for the peak value among each
level in the pyramid. This peak value contributes to rotation invariance and
high level of efficiency. As such, we decide to select key locations at the
maxima and minima of a difference of Gaussian function applied in scale space.
Maxima and minima of scale space of DoG are determined by comparing each pixel
in the pyramid to its neighbors. For instance, a pixel is compared to its 8
neighbors of the same level of the pyramid. If it is a maxima or minima at this
level, then the closest pixel location is calculated at the next lowest level
of the pyramid. The final step to construct SIFT keypoint is by implementing
orientation gradient equation as shown in formula (4). For each DoG image sample,
Lx, y, the gradient magnitude, m and orientation θ, is precompiled using
The construction of SIFT-C feature vector from keydescriptor will involve the
determination of orientation θ and magnitude m of every element in DoG
smoothing output. At each element in SIFT keydescriptors, we calculate the magnitude
of same angle direction and put into 1x128 keypoints. The implementation of
SIFT-C formula onto keydescriptor will produce values that can be used to classify
object. Here are the steps to construct SIFT-C values.
||Analyze of each k of SIFT keydescriptors by determine its magnitude at
every θ (128 bin of degrees).
||SIFT technique will produce kx128 keydescriptors which is constructed
by its magnitude m and orientation è. For every angle of θ,
sum its magnitude so that the keydescriptors are reduced [will reduce] to
||Implement to the narrow and broad weed and plot graph to visualize
the features of object.
||Finally, normalize [normalized] using [with] the maximum values of SIFT-C
The above algorithm can be described as (5):
Normalize the K(k,o)
As mentioned earlier, the 1x128 SIFT-C values are distinct and can be used
to represent the narrow and broad weed as unique feature vectors. This can be
clearly seen in the SIFT-C graph plot of Fig. 1. In order
to capture the differences, we used the statistical measurement method that
determines the Geometric Mean (GM) and Harmonic Mean (HM) of the data. The geometric
and the harmonic mean is
The statistics measurements above have been used as features to the input vector
for linear classifier design.
RESULTS AND DISCUSSION
A set of 1000 sample images consists of narrow and broad weed has been processed
using SIFT-C and obtained the keydescriptors. Figure 2 shows
the weed images and its SIFT-C keydescriptors plot graph. From the Fig.
2, we can see the SIFT-C values of broad and narrow that have different
tabulation, with maximum peak for broad graph occurred at SIFT-C keypoints axes
40-50 and 70-80 while for narrow, its maximum peak values occurred a few time
along the SIFT-C keypoints axes. Features for both types of weed have been extracted
by measuring its harmonic mean (7) and geometric means (8), respectively. To
prove the robustness of SIFT-C; we also test the technique onto image that has
a small weed distribution. The sample image of weed with small weed distribution
can be seen in Fig. 2. Normally, any changes on the content
of images will impinge on the accuracies of the processing technique. In our
observation (Fig. 3), using the SIFT-C technique indicates
that the pattern of graph for narrow and broad have very minimal changes as
compared to the normal image in terms of their pattern and values. The maximum
peak values of broad are still in the range of 40-50 and 70-80 while for narrow,
its maximum peak values occurred a few time along the SIFT-C keypoints axes.
Another sample of problem image that has been tested is an image that has been
enlarged to 50% of its original size of 240x320 pixels. This can be interpreted
as camera movement in a real time application. Resizing from original size of
image will automatically change the structure of the pixel values. This can
be seen when we compare the quality of the original and resized image. Usually,
any changes in size of an image will affect the final result of the classification
technique. Figure 4 shows the samples of problem images with
different resolution (zooming) by resize the original with 50% of its size.
In Fig. 5, original image has been adjusted by increasing
the brightness with different constant factor.
||Plot graph of 1x128 SIFT-C values of narrow and broad weed
||SIFT-C plot for image with little weed distribution (a) Broad and (b)
||Sample images of 50% zooming from original images (a) Broad and (b) Narrow
Constant factor of 1.5 and 2.5 will increase the brightness of original images;
meanwhile constant factor of 0.5 will give original images less brightness.
The sample image that used to verify the SIFT-C technique consists of all conditions
as discussed above. Figure 3-5 show the
types of sample images problem used for testing. These images represent the
various conditions that need to be faced and posed as a challenge to develop
a robust technique.
||Sample images of narrow and broad weed with different lighting effect.
(a) Narrow with constant lighting of 1.5 and 2.5 and (b) Broad with constant
lighting of 0.5 and 1.5
Using these types of images, the effectiveness of the SIFT-C technique can
be proven. The SIFT-C data has been analyzed using statistical measurement Geometrical
Mean (GM) and Harmonic Mean (HM) to extract its feature vectors. The feature
vectors which are obtained from (7) and (8) can be seen in Fig.
6. From the plotted graph, the feature vector of SIFT-C has been distributed
into two different clusters represent narrow and broad weed.
The feature vectors of raw images as shown in Fig. 6 measured
its statistical properties. The purpose is to see the patent of the values in
order to choose the best classifier tools for classification. It is clearly
seen in Table 1 and 2 that the statistic
values of both GM and HM are significantly justified in two different clusters.
For example, the mean of GM and HM for both types of weed are 0.32, 0.23 and
0.73 and 0.69. The other values of properties such as standard deviation, max.
and min. show that a linear classification can be chosen as a tool to classify
narrow and broad weed types.
||Feature vector of narrow and broad weed from original sample image
The same step was taken to measure the feature vectors to the problem images
which is the brightness has been increase and decrease by certain constant.
Figure 6-9 show the values of feature vector of their
respective brightness constant.
||Statistical properties of 1000 images of narrow and broad harmonic mean
||Statistical properties of 1000 images of narrow and broad geometrical
mean feature vector
||Classification efficiency of original sample image
||Percentage of classification result for different brightness
||Percentage of classification result for different zooming
||Feature vector of GM and HM for brightness constant 1.5
Increasing the brightness by scaling 1.5 doesnt affect the class of
feature vector. This can be seen in Fig. 6-9
where the values of GM and HM are still in two different clusters to represent
broad and narrow weed.
A similar case is depicted in the zoomed image (Fig. 10,
11). We enlarge the image to 70% from its original size and
the values of feature vectors are remain in the cluster of narrow and broad
||Feature vector of GM and HM for brightness constant 2.0
||Feature vector of GM and HM for brightness constant 0.5
||Feature vector of GM and HM for zooming 70%
This implies that the SIFT technique is robust towards different lightings
as well as unsteady camera movement.
||Feature vector of GM and HM for zooming 50%
Finally, the feature vectors obtained through HM and GM have been used as an
input to the classifier design. A simple classifier system as which was defined
in (9), (10) and (11) has been used and the equation obtained was f = -2.336559x1+44.193866x2-10.822155.
The linear equation f was used to test the classification efficiency of 1000
sample images. Table 3 shows the correct classification rate
of original sample weed image. It is found that the narrow and broad weeds are
well classified using SIFT-C with efficiency of 99.6 and 99.8%, respectively.
Table 4 and 5 show the testing results
of different image conditions that were manually adjusted from the original
image by varying the brightness of the images as well as resizing by 50 and
70%. The overall classification rate of brightness images with constant factor
0.5 remains unchanged. Further decrease to a factor of 0.2 causes the accuracy
to drop to 90.7%. This also happened to the image with constant factor of brightness
increase up to 2. The efficiency was dropped but still in the acceptable error.
The fundamental idea of image processing techniques is based on structural analysis of SIFT-C technique. Ishak et al. (2007) used a statistical analysis of histogram technique to classify the same weed in palm oil plantation. The result obtained from her studies showed the poor classification efficiency with the overall correct classification rate at 88.75%. Statistical approach of image processing technique is very sensitive to its pixels value as any changes of image due to noise such as lighting or bad resolution will automatically affect the final result. The use of statistical approach of SIFT-C technique was successful classified the type of weed as the efficiency is constantly above 90%.
The adoption of SIFT for object classification instead of object detection
was successful in classifying narrow and broad weed. Figure 6
shows the overall results of successful classification with different image
conditions. We test our image by manually adjusting the brightness to see the
robustness of SIFT-C technique. After increasing the brightness of the pixel
value by a factor of 1.5, the classification rate remains until the factor was
doubled. At this point, the classification rate decreases to 94 and 94.8% for
narrow and broad weed, respectively. We also test for different sizes of image
as this can be perceived as zooming or camera moving in real application. The
performance rate does not change much for the 70% decrease of the image. To
conclude, the SIFT-C has a robust characteristic and shows better performance
for narrow and broad weed classification.