A convenient and robust edge detection method based on ant colony optimization

Highlights

• A convenient and robust ACO-based edge detection method is proposed.

• A new heuristic function is devised to detect more edge points.

• A user-defined threshold in the pheromone update process is adopted.

• For ease of use, a group of suitable parameter values are provided.

Abstract

Edge detection is usually used as a preprocessing operation in many machine vision industrial applications. Recently, ant colony optimization (ACO) as a relatively new meta-heuristic approach has been used to tackle the edge detection problem. In this work, a convenient and robust method for edge detection based on ACO is proposed, which employs a new heuristic function, adopts a user-defined threshold in pheromone update process and provides a group of suitable parameter values. Experimental results clearly demonstrated the effectiveness of the proposed method, and at the same time, in the presence of noise, the proposed approach outperforms other two ACO-based edge detection techniques and four conventional edge detectors.

Introduction

Edge detection is an essential preprocessing operation in many machine vision industrial applications [1], [2], [3], such as shape recognition, 3D reconstruction, defect detection on mechanical parts, etc. Edges are sets of pixels in the image regions with sharp intensity changes and correspond to visible contour features of objects in an image. Normally, edge detection is a process that inputs a grayscale image and then outputs a binary image to indicate the edges of objects.

Many edge detection methods have been proposed in the last decades. Most of them are based on digital differential methods [4], such as Sobel, Roberts, Laplacian operators, and so on. However these algorithms are quite sensitive to noise, in order to suppress noise, Marr et al. [5] and Canny [6] applied the Gaussian pre-smoother to the image before detecting edges. Unfortunately, this procedure would blur edges while removing noises in the image. Consequently, these approaches would sacrifice the locating accuracy of the detected edges to a certain extent [7], as shown in Fig. 1. To overcome this limitation, edge detection could be formulated as an optimization problem. Ant colony optimization (ACO) [8] as a relatively new optimization approach has been used for edge detection, which could be classified into two categories: direct edge detection [9], [10], [11], [12], [13], [14], [15], [16], [17], [18] and broken edge compensation [19], [20]. In this work, the proposed method is based on the former, because the latter is used just as a complementary tool to other edge detectors.

ACO as a swarm intelligence approach has been adopted to directly detect image edges by Zhuang [9], [10] since 2004. He used the Ant Colony System (ACS) to build the perceptual graph of images for extracting edge features. Unfortunately, his two methods are only capable of detecting simple edges. After that, Nezamabadi-Pour et al. [11] exploited the Ant System (AS) and applied the directed graph to detect edges. Even though they derived the relationship between image area and the number of ants, they did not use other information about the image for more parameter setting. In practice, the more the adaptive parameters are provided, the more conveniently the proposed method can be used. In this work, we employed three adaptive parameters. Tian et al. [12] also adopted the ACS and proposed the method of computing adaptive threshold to tackle the edge detection problem. Similarly, Jevtic et al. [13] first used multiscale adaptive gain for image contrast enhancement, then applied the ACS to detect image edges. Since then many adaptive thresholding methods [14], [15], [16], [17], [18] have been presented. But in fact, the ACO meta-heuristic approach as a swarm intelligence technique is inherently adaptive, because this technique is a collective behavior of decentralized, self-organized agents in a swarm. Therefore, in this work, a user-defined threshold in the pheromone update process is adopted to take advantage of this feature of ACO, and noises can be suppressed effectively by adjusting the user-defined threshold.

In addition, the heuristic information matrices applied to ACO-based edge detection could be generally divided into two types: one is proposed by Nezamabadi-Pour et al., which was used in literatures [11], [13], [15], [16]; the other adopted in [12], [14], [17], [18] was devised by Tian et al. On the basis of the two types of heuristic information matrices, a new heuristic information matrix (i.e. Eq. (5) in Section 3) is proposed in this work to improve the gradient response on the edge.

The remaining sections of this paper are organized as follows. First, the main differences between AS and ACS and the implementation of AS algorithm are briefly described in Section 2. Then, the proposed AS-based method for detecting image edges is presented in Section 3. Next, the parameter setting of the proposed approach and the experimental results of performance comparison are given in Section 4. Finally, the conclusions are drawn in Section 5.