Using an improved colonial competition algorithm for image segmentation

Dissertation for Master's Degree

Computer Engineering - Artificial Intelligence

Abstract

Image segmentation is a fundamental process in many applications of image processing and machine vision, which can be considered as the first low-level processing step in digital image processing. Image segmentation has various applications such as medical image processing, face recognition, traffic control systems, etc. Due to the importance of digital image segmentation, several methods have been proposed for this purpose, which are divided into two general categories of area-based methods such as image pixel clustering and edge detection-based methods. Most image clustering methods categorize pixels only based on their brightness or color information and do not use any neighborhood or location information of pixels in the image clustering process, which reduces the accuracy and quality of segmentation. Considering the importance of using the spatial information of pixels in order to improve the quality of image segmentation, using the information of neighboring pixels in the large neighborhood window improves the quality of segmentation. Considering that clustering is considered to be part of non-deterministic-hard polynomial problems, in this research the idea of ??combining k-means clustering algorithm and improved colonial competition algorithm is proposed to solve this problem. Also, before applying the combined algorithm, a new image is created using the non-local information of the pixels, and then the combined algorithm is used to cluster the pixels of the new image. By comparing the results of applying the mentioned method on different images with other methods, we came to the conclusion that the accuracy of segmentation of most of the images with the proposed method is higher than other algorithms in this field. Keywords: image segmentation, clustering, improved colonial competition algorithm and non-local information Chapter 1 Introduction Image processing today more Digital image processing is a branch of computer science that deals with digital signal processing that represents images captured by a digital camera or scanned by a scanner. Image processing has two branches: image enhancement [2] and machine vision [3]. Image enhancement is methods such as blurring and contrast enhancement to improve the visual quality of images and ensure they are displayed correctly in the destination environment (such as a printer or computer monitor), while machine vision deals with methods that can be used to understand the meaning and content of images to be used in tasks such as robotics. Image segmentation is one of the most important basic steps in digital image processing. Image segmentation is the separation of image pixels into separate areas that are the same or correlated as much as possible in terms of features such as brightness, texture, or color. Zoning images in many processing tasks on images, such as image therapy, machine vision, image compression, objectology, is a necessary and important requirement to start processing on the desired object or texture. 

     Image segmentation is done in different ways, which can generally be divided into two categories: classical and morphological. In classical and traditional methods, changes in brightness are used in order to extract the edges and local characteristics of the desired objects. Another type of classical algorithms are methods based on statistical algorithms in which segmentation is done based on the distribution of pixels and finding a suitable threshold. Since the images often have noise, distortion [4], occlusion, exposure to illumination and so on, these methods are unusable for many applications.  Newer methods that are used today, use classification (clustering), to zone and divide the image. These algorithms are Fuzzy C-means and K-means clustering algorithms, neural network algorithms such as simple competitive training [5], simple Bayes training tree, etc. Although these methods have good detection accuracy, they are highly dependent on initialization (in K-means initialization of cluster centers and training rate in neural networks) and the algorithm must be applied to the image many times to get the optimal solution.However, convergence in these methods is not always guaranteed and in some cases they are trapped in local optima. Finding the optimal centers of image clusters is one of the hard non-polynomial problems. On the other hand, in most methods, clustering of pixels is done based on features such as color or brightness, and no spatial information or neighborhood of pixels is used, which makes these methods ineffective in segmenting noisy images. Also, based on the principle that the neighboring pixels in the image are interconnected and considering that most of the pixel clustering algorithms segment the pixels of the image without considering the similarity between the neighboring pixels, the image becomes so-called over-segmentation [6] (it is divided into a large number of areas). Image segmentation using clustering and using pixel neighborhood information has received the attention of researchers in recent years. Ahmad and his colleagues introduced the information of local brightness intensity by modifying the objective function of the FCM algorithm [7] for image segmentation so that the labeling of pixels was done under the influence of their local neighborhood [1]. Chen and Zhang introduced two new versions of the FCM algorithm in which the neighborhood term was calculated before applying fuzzy clustering [2]. Zillagi and his colleagues introduced an improved FCM in order to accelerate image segmentation, a linear weighted sum image was calculated using the original image, and then the FCM clustering algorithm was applied to the histogram [8] of the newly created image [3]. The FGFCM algorithm [9] was proposed by Kai and his colleagues [4]. The performance of this method was based on the creation of a new image using a similarity measure that combines spatial information and local light intensity information. Halder and his colleagues presented an evolutionary approach for unsupervised image segmentation, the purpose of which was to cluster pixels based on information of brightness intensity and pixel neighborhood relations, using genetic algorithm [5]. The idea of ??using the non-local information of pixels, which was based on the non-local average filter, was proposed by Zhao and his colleagues [6]. In this thesis, the K-means clustering algorithm, which is one of the most common clustering methods, is combined with the improved colonial competition algorithm. The objective function of the K-means algorithm is used in the improved colonial competition algorithm in order to find the optimal centers of image pixel data clusters, and a pre-processing step is applied to the input image in order to use the local and non-local information of the pixels before running the colonial competition algorithm. The idea of ??using non-local information is taken from the non-local average filter that was proposed to reduce the effect of Gaussian noise [10]. To see the results, we applied the proposed algorithm on synthetic images degraded with Gaussian noise and also on natural images. The continuation of the thesis is organized as follows, in the second chapter, the description of the problem states that image segmentation is introduced and the input and output of the problem and the purpose of performing segmentation on the images are examined. In the third chapter, the concepts of pies are discussed and the K-means clustering algorithm and the improved colonial competition are described, and then we get to know some concepts of image processing and machine vision. In the fourth chapter, based on the investigations, we will review the past works in the field of image segmentation and noisy image segmentation. In the fifth chapter, the details of the proposed method including the improvement of the colonial competition algorithm and its combination with the K-means algorithm, the use of non-local information of pixels and the improvement provided for its use in image segmentation are reviewed. In the sixth chapter, the results of segmenting different images with the proposed algorithm and comparing it with other methods will be analyzed and reviewed, and finally, in the seventh chapter, we will conclude and examine the advantages and disadvantages of the proposed method and future solutions.