This research work presents an improved edge detection algorithm using particle swarm optimization based on vector order statistics. The proposed algorithm was implemented using MATLAB 2013 script. The algorithm addressed the performance of edge detection in images, with a view to minimizing broken, false and thick edges whilst reducing the presence of noise as well as computational time. A collection scheme based on step and ramp edges was developed for the edge detection algorithm, which explores a larger area in the images in order to reduce false and broken edges. The efficiency of this algorithm was tested on two Berkeley benchmark images in clean and noisy environments with a view to comparing results, both visually and quantitatively, with those obtained using proven edge detection algorithms such as the Sobel, Prewitt, Roberts, Laplacian and Canny edge detection algorithms. The algorithm was also applied to facial and remotely sensed images with a view to testing the algorithm on real life images. The Pratt Figure of Merit (PFOM) was used as a quantitative comparison between the developed algorithm and the proven edge detection algorithms. The benchmark value for the PFOM is between 0-1, which shows efficient detection of edges as the value tends towards 1. The quantitative results obtained using PFOM on the test images in clean environment for the Sobel, Prewitt, Roberts, Laplacian, Canny and the proposed edge detection algorithms are 0.4209, 0.4195, 0.4181, 0.7048, 0.8421 and 0.8480, respectively. This showed that the proposed algorithm detected more edges in clean environment as the value obtained is nearest to 1. The PFOM on the test images in noisy environment for the Sobel, Prewitt, Roberts, Laplacian, Canny and the proposed edge detection algorithms are 0.4191, 0.4191, 0.2807, 0.2811, 0.5606 and 0.8458 respectively. This showed that the proposed algorithm detected more edges in noisy environment as the value obtained is nearest to 1. The proposed algorithm achieved a PeakSignal-to Noise Ratio (PSNR) of 57.7320dB in environment containing ≤ 33% of noise level. This result signifies 3% improvement in detection of edges in noisy environment as compared with the proven traditional edge detection algorithms which achieved an average PSNR of 22-35dB.
TABLE OF CONTENTS
TABLE OF CONTENTS vii
LIST OF FIGURES ix
LIST OF TABLES xi
LIST OF ABBREVIATIONS xii
CHAPTER ONE: INTRODUCTION
1.1 Background 1
1.2 Statement of Problem 3
1.3 Aim and Objectives 4
1.4 Methodology 4
1.5 Significant Contributions 5
1.6 Dissertation Organization 5
CHAPTER TWO: LITERATURE REVIEW
2.1 Introduction 6
2.2 Review of Fundamental Concepts 6
2.2.1 Image processing 6
2.2.2 Gray-scale image 7
2.2.3 Coloured images 8
2.2.4 Image filtering 10
2.2.5 Image convolution 12
2.2.6 Edge detection 13
2.2.7 Review of existing edge detection algorithms 17
2.2.8 Image noise 24
2.2.9 Particle swarm optimization (PSO) 25
2.2.10 Setting particle swarm optimization parameters 29
2.2.11 Vector order statistics 31
2.2.12 Calculating euclidean distance between two pixels 33
2.2.13 Edge intensity 34
2.2.14 Benchmark test images used 35
2.2.15 Real life images used 36
2.2.16 Pratt figure of merit (PFOM) 40
2.3 REVIEW OF SIMILAR WORKS 41
CHAPTER THREE: MATERIALS AND METHODS
3.1 Introduction 50
3.2 Methodology 50
3.3 Profile Modelling of Edge Intensity 51
3.4 Collection Scheme 51
3.5 The Proposed Edge Detection Algorithm 55
3.5.1 Particle swarm optimization algorithm 55
3.5.2 Algorithm for the vector order statistics 57
CHAPTER FOUR: RESULTS AND DISCUSSIONS
4.1 Introduction 61
4.2 Analysis of the Proposed Edge Detection Algorithm 61
4.1.1 Computation time 65
CHAPTER FIVE: CONCLUSION AND RECOMMENDATIONS
5.1 Introduction 74
5.2 Summary 75
5.3 Conclusion 76
5.4 Limitations 75
5.5 Recommendations and Further Work 76
Edge detection can be defined as the process of identifying set of connected pixels that forms a boundary between two disjoint regions (Gang et al., 2008). It can also be defined as the process of locating and identifying sharp discontinuities in images (Rashmi et al., 2013). It is mostly used in image analysis to preserve image features and partition images into regions of interest. The discontinuities in these images can be caused by (Ghasemi et al., 2011):
i. Discontinuity in depth and/or surface colour and texture.
ii. Reflection of light, Shadows and Illumination.
Edge detection is an image segmentation technique in which images are partitioned into meaningful regions of interest. Some of the practical applications of edge detection algorithms are in face and finger print recognition, location of objects in satellite images, medical images, and computer aided surgery or diagnosis amongst others (Rashmi et al., 2013). One of the most important challenges of edge detection algorithm is to detect the edges in noisy images. Many traditional edge detection algorithms have been developed to overcome noise such as Sobel, Prewitt, Roberts and Gradient based edge detection algorithms etc. (Rashmi et al., 2013). These traditional edge detection algorithms are very fast but they cannot perform well on noisy images. Hence, the significant problem of these edge detection algorithms are displacement, removed edges, false and broken edges(Maini & Aggarwal, 2011). Noise phenomenon is an obstacle in detection of continuous edges as it causes some variation of pixel intensities, thus reducing the performance of an edge detection algorithm in noisy images (Setayesh et al., 2013). It also leads to unclear and displaced edges (Chaudhary & Gulati, 2013). Many edge detection
algorithms have been developed in the literature over the past years to improve precision of recognized edges. However, they still suffer from producing broken edges and false edges due to noise effect (Maini & Aggarwal, 2011). Therefore, an improved edge detection algorithm is required to detect edges with greater continuity in noisy images in order to reduce the shortcomings of traditional edge detection algorithms. In the field of image processing, there exists basically two types of images which are the gray scale and the coloured images. Numerous researchers have developed edge detection algorithms for gray scale images in the past. But in recent times, with improvement in computer capabilities and the increased applications of coloured images there is need to develop an effective edge detection algorithm for coloured images (Haque & Aljahdali, 2013). Some of the applications of the edge detection algorithms are in image segmentation, image compression, face recognition, computer vision, computer surveillance, medical diagnosis, image encryption/communication multimedia and remotely sensed images, amongst others(Vijayarani & Vinupriya, 2013). In the areas of medical diagnosis, satellite images, face recognition, and computer surveillance, representation of images by its edges reduced the amount of data required to be stored whilst retaining useful information in the image.
Most edge detection algorithms process a single pixel on an image at a time and calculate a value which shows the edge magnitude of the pixel, and the edge orientation. Then, a thresholding technique is utilized to recognize if a pixel is an edge or not (Rai & Dutta, 2013) and the result will be a binary image which indicates the location of all existing edges on the original image. The edges detected by these algorithms are not usually linked and there is no relation among the edge pixels. To solve this problem, most edge detection algorithms utilize a linking technique such as the Hough transform, sequential edge linking etc. However, the linking process in such
techniques is not good except for the edges on simple shapes such as circles or lines(Setayesh et al., 2013). Most of the edge detection algorithms weredeveloped to perform on gray level images as compared with coloured images. This is because edge detection in coloured images is a far more challenging task, and the criteria for a good edge detection algorithm are (Maini & Aggarwal, 2011):
i. The optimal detection must minimize the probability of false positives (detecting spurious edges caused by noise), as well as that of false negatives (missing real edges).
ii. The edges detected must be as close as possible to the true edges.
1.2 Statement of Problem
The traditional edge detection algorithms use limited or small area to detect a pixel as an edge. Edge detection is very crucial in image processing (Ghasemi et al., 2011) and size of the area being considered has strong influence on the accuracy of the detection (Setayesh et al., 2013). The larger the area, the less the sensitivity to noise, but at the same time, the localization accuracy is lower. In order to increase the localization accuracy of the algorithm, there is the need to consider all the edge patterns. However, this increases the computation time and produces broken and false edges. In this research work, a particle swarm optimization edge detection algorithm for coloured images based on vector order statistics is proposed in order to reduce false and broken edges as well as computational time by exploring a larger area in the noisy images
1.3 Aim and Objectives
The aim of this research work is to develop a particle swarm optimization (PSO) edge detection algorithm for noisy coloured images based on vector order statistics with a view to reducing false and broken edges as well as computational time by exploring a larger area in the noisy images. The objectives are as follows:
i. Development ofacollection scheme for set of pixels in coloured images with a view to reducing false and broken edges in the image.
ii. Development of a particle swarm optimization edge detection algorithm for noisy coloured images based on vector order statistics.
iii. Validation of the proposed algorithm in (ii) and comparison with the traditional edge detection algorithms using Pratt Figure of Merit (PFOM)
The following methodology was adopted in carrying out this research:-
i. Developingacollection scheme to detect edges in coloured images.
ii. Exploring a larger area and examine normally occurring edge patterns in order to increase the localization accuracy of edge detection and determine which edge pixels should be discarded as noise and which should be retained.
iii. Extracting the global structure of edges in order to detect the edges with greater continuity. Hence, determine the exact location of an edge.
iv. Testing the proposed algorithm in MATLAB 2013b image processing toolbox on two coloured images obtained from the Berkeley benchmark image database, remotely sensed image generated using Google earth software and real facial images.
v. Validation and testing of the proposed edge detection algorithm
The existing traditional edge detection algorithms use a single pixel on an image at a time to calculate a value which shows the edge magnitude of the pixel, and the edge orientation. However, this leads to false and broken edges in the generated output edge map. Therefore, the significant contributions of this research work are itemized as follows:- i Development of a scheme for collection of pixels based on Step and Ramp edges with a view to reducing false and broken edges that exists when generating the output edge map.
ii Application of Vector order Statistics based on collection of pixels for edge detection in noisy coloured images. An improved Pratt Figure of Merit (PFOM) value of 0.09% and 33.7% were obtainedin clean and noisy environments, as compared to the best amongst the existing traditional edge detection algorithms.
iii The proposed edge detection algorithm produced thin and continuous edges in noisy environment and achieved a PSNR of 57.732dB. This represented a 3% improvement in detection of edges in noisy environments as compared with other proven techniques such as the Sobel, Prewitt, Roberts, Laplacian and Canny.
1.6 Dissertation Organization
The general introduction has been presented in chapter one. The rest of the chapters are presented as follows: a detailed review of the fundamental concepts of image processing, edge detection as well as a review of similar research works is presentedin chapter two. Profile modelling of edge intensity, mathematical equations and formulation of the problem are
presented in chapter three. Analysis and discussions of the results are presented in chapter four. Summary, conclusions limitations and recommendations are presented in Chapter Five.