APDRS DESIGN AND METHODOLOGY

In this chapter we shall attempt to solve the full pest problem by combining two existing techniques, which are capable of achieving invariance to several of the possible variations of the target object.

The automatic plant pest detection and recognition algorithm presented in this thesis uses k-mean clustering (see section 3.2) and the Correspondence filter (see section 3.3.1). As a test bed we investigate a collection of plant pests, which were obtained from different online resources, among which are: (ALABAMA Nursery and Landscape), (Iannotti, 2012) and (Hamilton, 2012), (Yarham, 2013), (Cossey, 2007). The algorithms have been tested on more than ten crop or plant pests as shown in Fig.7.2

Owing to the risks crop pests impose on livestock and mammals, in a bid to control pest invasion and reduce the risks from chemical pesticide, pest detection and recognition algorithms were proposed by (Faithpraise et al., 2013a). The algorithms were divided into two operational phases, the detection phase and the recognition phase because of the complex nature of the object pest. The proposed pest detection and recognition algorithms are explained in the flow chart of Fig. 7.3.

Grape root borer Green peach aphids Japanese beetles Fallworm Caterpillar Spittlebug Tortoiseshell Butterflies

Alfalfa aphids Leaf roller katydid _Armyworm Helicoverpa

larva Cherry leaf _hoper

DESIGN AND DEVELOPMENT METHODOLOGY 2014

Faithpraise Fina Otosi CHAPTER 7 Page 119

Moth _{Scale Potato beetle Brown snail White snail Slug}

Spider

Odonata Anopheles

mosquito Odonata eggs Elephant _mosquito

Pea Aphid

Mosquito Spodoptera Odonata

nymph

Ladybug

Larvae Cutworm Elephant mosquito larvae

Figure 7. 2 Crop Pests & NBI in their habitats

7.2.1 DETECTION SYSTEM

The APDR algorithm starts with the Pest Surveillance system acquiring the digital images from the environment and loading them as the filter construction images; following acquisition the colour transformation structure for the RGB plant pest images was created. A device-independent, L*a*b* (abbreviation for the CIE 1976 (L*, a*, b*) (or CIELAB) which consists of luminosity L*, chromaticity layer a* and chromaticity layer b*, which houses all the colour (red-green and blue-yellow axis) information, colour space transformation structure was applied; the k-means clustering technique was then used to segment the plant pest images. Now the object(s) or pest segmentation process begins by identifying the

principally green colour pixels based on using a varying threshold, mainly all the red, green and blue colour components of the pixel are assigned a zero value so that all the pixel intensities that are less than the threshold value will be deleted. This technique is applied in as much as we assume that, these pixels are not useful for the pest identification. All the pixels on the boundaries of the object cluster and all the colour pigments, which are designated zero, were deleted totally and the segmented pest result is shown in chapter 8, Fig. 8.1. The same steps were repeated for each pest image in the data set.

Faithpraise Fina Otosi CHAPTER 7 Page 120

Figure 7. 3 Proposed plant pest detection and recognition process flow chart

The problem of identifying crop pest species despite distortion in position, orientation and scale when occluded or hidden in its cluttered habitat (cluttered background refers to plants or leaf) is a demanding problem that requires an immediate solution to offer the right control measure. Detecting the pest images from the clustered environment as shown in Fig. 8.1, of chapter 8, is the first step to establish the presence of insect objects on the crop. To confirm the presence of pest species, the position and the population density on the habitat, a recognition system is required.

7.2.2 RECOGNITION SYSTEM

The second stage of the APDR algorithm began by angle multiplexing the pest detected from the detection algorithm into the correspondence filter at 50 increments. The recognition system is expected to correctly detect the pest images despite variation in its position or

DESIGN AND DEVELOPMENT METHODOLOGY 2014

Faithpraise Fina Otosi CHAPTER 7 Page 121

orientation. To create invariance to orientation changes of the pest, the correspondence filter was created using a training image set consisting of the expected range of rotations taken at small intervals of viewing angle as shown in chapter 8, Fig. 8.2. The correspondence filter uses a small size moving window system to scan the input image in one pixel increments in order to test every location of the input image for the presence of a target. We discovered that there was some tolerance to the position of the filter centre relative to a target centre. The moving window was then increased to two pixels and there was an insignificant reduction in the detection of the target object.

The correspondence filter has been tested precisely for its detection and recognition accuracy especially for its distortion invariance ability. The sub-database is made up of 10 training images rotated in plane by 5o increments between each training image of the pest. To optimise the recognition performance of the correspondence filter, suitable values of the tuning parameters alpha, beta and gamma had to be selected. The values of these parameters were picked by observing the impact of different values for a number of trials until the most appropriate values that yield the best recognition ability was established. For these correspondence filters the tuning parameters chosen were = 0.0000009; = 0.45 and = 0.1. The value was set so low to produce sharp peaks while still being tolerant to noise in the habitat. was set at a large value to produce sharp peaks with good distortion tolerance. value was made low to maximise the discrimination ability of the filter while still producing good distortion tolerance.