Robust shape-preserving contour tracing with synchronous redundancy pruning

Abstract

This paper presents a concise, robust shape-preserving contour tracing scheme for efficiently extracting a highly representative one-pixel-width closed contour for the profiles of heterogeneous objects from digitally processed images regardless of noise existence. The tracing mechanism comprises two synchronous-running modules, namely a progressive boundary linking (PBL) module and a synchronous redundancy pruning (SRP) module. PBL applies concise pixel-based decision-making criteria to select and links suitable pixels belonging to the object’s contour in real-time. At the same time, SRP simultaneously removes random noises and prunes perceptually trivial branches/cracks, which are regarded as commonly-acknowledged redundancies, from the constructing contour curve. Thus, the relatively meaningful details of original object profile can be preserved on the extracted contour. The performances of proposed contour tracing method in extracting object contours from clean and noisy images are both qualitatively and quantitatively compared with that of a conventional digital signal processing (DSP) technique and mathematical morphology boundary extraction method, and integration of the mathematical morphology noise-cleaning process and conventional DSP technique, respectively. Overall, the simulation results indicate that the one-pixel-width closed contours extracted by the proposed mechanism have a greater degree of precision and noise suppression than those extracted by the conventional DSP method, the mathematical morphology boundary extraction method, and conventional DSP method following morphological cleaning process. Therefore, the proposed contour tracing mechanism could offer a viable efficient contour-extraction tool in supporting some well-known image processing applications such as object shape-coding, pattern-recognition and posture change detection, and so forth.

Introduction

Border tracing is a technique for accurately extracting the contours of an object from a digitally-processed image. Such techniques are widely applied for preprocessing purposes in many image/video multimedia applications, including pattern recognition based on shape features, target tracking using object gestures (Ueda and Suzuki, 1993, Li and Narayanan, 2003), shape coding (MPEG-4 Video Verification Model v18.0, 2001), and so forth. However, modern multimedia applications commonly involve the use of image/video segmentation, compression and even simple digitization routines, which inevitably introduce noise (i.e. distortion) into the original images. One of the most serious effects of this noise is the contamination of the boundaries between neighboring image objects such that they are no longer distinct. As a result, the quality of subsequent image/video automatic recognition and shape-coding applications, in which the object boundary information is of fundamental importance, is inevitably degraded. Generally speaking, the image contamination caused by random white noise can be suppressed linearly through low-pass filtering. However, fragmentation of the object border caused by burst errors is not so easily compensated using unsupervised methods. In general, redundancies of continuous area in an image do not only result from burst errors, but also from some image processing operations. For example, object boundary fragmentation is commonly caused by image segmentation (Beucher and Lantuéjoul, 1979, Fan et al., 2001), particularly when the colors or gray-levels in the original image vary progressively across the boundaries between neighboring objects. When applying common contour extraction methods to noisy object boundaries, it is generally found that morphology-based methods (Giardina and Dougherty, 1988, Meyer and Beucher, 1990, Gonzalez and Wood, 2002) provide more accurate results than those obtained using edge-detection schemes (Gonzalez and Wood, 2002). However, visible distortion of the object profile is often apparent when morphology methods based on nonlinear filters (conventionally referred to as structuring elements (SE’s)) are applied (Costa and Cesar, 2001). Such methods not only suppress undesirable noise within the image, but also inadvertently eliminate nontrivial details such as textural branches, cracks, shape contour features, and so forth. Crucially, the resulting contour distortions may have a seriously detrimental effect upon the performance of subsequent pattern recognition applications, such as those used for medical diagnosis, remote sensing or detection, and so forth. Therefore, there is a practical need to develop robust border tracing schemes which effectively suppress the noise content of an image, while simultaneously retaining the fine details of the object contours.

Accordingly, the current study develops a robust contour-tracing scheme with the ability to extract one-pixel-width contours of image objects which faithfully preserve the detailed characteristics of the original object profile, while simultaneously eliminating trivial branches/cracks. Importantly, the proposed method is capable of tracing the contours not only of clean, well-defined image objects, but also those of objects which are subject to noise distortion as a result of region-based segmentation, noisy transmission, faulty de-compression, and so forth. In developing the contour tracing scheme, a fundamental assumption is made that the object pixels have already been separated from the background pixels via some ordinary image segmentation processing, where a region-based technique is usually adopted. In other words, the contour tracing scheme is designed for processing the images within which heterogeneous objects have already been distinguished from each other.

The proposed method comprises two discrete but synchronously integrated mechanisms, namely the progressive boundary linking (PBL) mechanism and the synchronous redundancy pruning (SRP) mechanism. The PBL mechanism is basically designed to link successive pixels along the object boundary so as to form a closed contour of the object profile, while the SRP routine removes fragments, random noise and trivial branches/cracks from the object border in order to improve the definition of the constructed contour. The integrated PBL-SRP contour tracing method, designated as PSCTM hereafter, is designed to extract a one-pixel-width closed contour of an entire object/demarcation boundary between adjacent regions from a noisy image without the visible deformation associated with existing contour tracing methods such as morphological boundary extraction (Giardina and Dougherty, 1988). Importantly, the functionality of the SRP mechanism is quite different from that of conventional morphological noise cleaning schemes (Gonzalez and Wood, 2002, Giardina and Dougherty, 1988, Peters, 1995), which inevitably smooth not only the noise in an image, but also the sharp regions of the object border, and therefore risk losing the unique characteristics of the contour profile. In PSCTM, the SRP module is designed to improve the identification of the extracted contour constructed by PBL by eliminating perceptual redundancies along tracing of the object profile, such as those caused by random snow noise for example. In other words, SRP removes redundancies from the image by pruning rather than performing a smoothing operation, such as morphological cleaning does, and subsequently preserves the distinctive characteristics of the object profile at the same time.