Fast and accuracy extraction of infrared target based on Markov random field

Abstract

Infrared images are characterized by small signal-to-noise ratio (SNR) and low contrast thus making it much difficult to achieve accurate infrared target extraction. This paper proposes a fast and accurate segmentation approach to extract targets from an infrared image. First, the regions of interests (ROIs) which contain the entire targets region and a little background region are detected based on the variance weighted information entropy feature. Second, the infrared image is modeled by Gaussian Markov random field (MRF), and the ROIs are used as the target regions while the remaining region as the background to perform the initial segmentation. Finally, by searching solution space within the ROIs, the targets are accurately extracted by the energy minimization using the iterated condition mode (ICM) based on the fact that targets can only exist in ROIs. Because the iterated segmentation results are updated within the ROIs only, this coarse-to-fine extraction method can greatly accelerate the convergence speed and efficiently reduce the interference of the background clutter and noise. Experimental results of the real infrared images demonstrate that the proposed method can extract single and multiple infrared targets accurately and rapidly.

Introduction

Infrared images provide useful information for many applications, such as military navigation, automatic target recognition and tracking [1]. Extraction of infrared targets from an infrared image is particularly important, as it plays an important role in automatic infrared target recognition and infrared target tracking systems. As we know, many infrared imaging systems produce images with small signal-to-noise ratio (SNR) and low contrast, which reduces the detectability of targets and impedes further investigation of infrared images. It is a challenging task to extract the targets from an infrared image accurately and rapidly.

So far, numerous infrared image segmentation methods for extracting the target regions from background have been proposed, e.g., the method based on thresholding [2], [3], [4], [5], [6], the method based on clustering [7], the method based on edge detection [8], [9] and the method based on splitting and merging [10], the method based on level set [11] and the method based on graph cut [12]. Thresholding-based method is widely used in infrared target extraction, such as one-dimensional (1-D) Otsu [3], two-dimensional (2-D) Otsu [5], 2-D entropy [6], etc. However, the thresholding-based methods cannot achieve satisfactory results for the infrared images with complex background and low SNR, since they ignore the spatial relationship of the image pixels which leads to increased sensitivity to image noise, increasing the error rate. They are likely to classify some of the background into foreground (targets) for the infrared images.

In the past 20 years, Markov random field (MRF) has been developed as a powerful tool for image segmentation because it can capture the spatial constraints among pixels and merge the context information into segmentation process. There are various MRF-based segmentation models that have been developed. Cohen and Cooper [13] proposed a doubly MRF model for segmenting range images and natural scenes. Melas and Wilson [14] applied the double MRF model with modification to segment satellite images. Won and Derin [15] developed a hierarchical MRF model for segmenting noisy and textured images. Scherrer et al. [16] constructed the distributed local MRF models for tissue and structure segmentation of medical images. Poggi et al. [17] did research on supervised segmentation of remote sensing images based on a tree-structured MRF model. Panjwani and Healey [18] extended the coupled MRF models to segment textured color images. Deng and Clausi [19] proposed a segmentation of SAR sea-ice image using a simple MRF with a variable weighting parameter. Xu and Luo [20] proposed a new nonhomogeneous MRF model based on fuzzy membership for brain MRI segmentation. However, the application of MRF models to segment infrared imagery is not commonly represented in the research literature. It is mainly because for the infrared images with complex background and low SNR, MRF-based model is likely to classify some of the background into foreground. In addition, the classic MRF model is complicate and has low convergence speed, which impedes its application for infrared image segmentation.

The goal of this paper is to propose a fast and accurate segmentation method for infrared target extraction based on MRF. First, the regions of interests (ROIs) which include the entire targets regions and some background regions are detected by using the variance weighted information entropy (WIE) [21]. Then, the infrared image is modeled with MRF, and the ROIs are used as the target regions and the remaining region as background to do the initial segmentation. Finally, by searching solution space in the ROIs based on the fact that targets can only exist within the ROIs, the targets are accurately extracted by the energy minimization. This coarse-to-fine extraction method can reduce the interference of the background clutter and noise in a large degree and decrease the search space greatly.

The rest of the paper is organized as follows. Section 2 introduces ROIs detection based on the variance weighted information entropy. Section 3 first describes the MAP–MRF segmentation framework, and then introduces the accurate target extraction based on MRF. Section 4 is devoted to experimental results for infrared target extraction with the proposed and other techniques. Conclusions are drawn in Section 5.