Segmentation of Images with Insufficient Dynamical Range

In this paper a method is presented to automatically detect objects from images with insufficient contrast. Using the edge map by the Canny edge detector as a reference, the corresponding edge structures in the edge map by the zero-crossing edge detector are firstly determined. Then meaningful structures are differentiated with respect to those randomly distributed edge segments and are recovered from the zero-crossing edge map. The method is validated in a realistic vision system and compared favorably with existing methods. 1. General Instructions As a fundamental issue in computer vision and image processing, object detection aims to find semantic objects in digital images or videos. The detection can be supervised with an object model built upon a set of training data, like the Viola-Jones object detector. What is of interest here is the unsupervised object detection, which usually starts from a basic assumption: pixels associated with the same object will share similar properties (like intensity or standard deviation in a neighborhood) and pixels associated with different objects will exhibit different properties. There are two ways to apply this assumption to object detection by emphasizing the first part or the second part of the statement. An object can be detected through detecting all pixels with similar properties. In this way, object and background can be differentiated, and a partition of the image can be derived, which is named region-based image segmentation. Alternatively, one can detect the object through locating the boundaries since image properties will change at object boundary pixels. This approach is called edge detection. There have been a large number of techniques that had been proposed since 1960s to address the issue. Earlier developments treat the pixels separately, but efforts after 1980s tend to jointly consider the spatial relationship in order to yield a smooth segmentation map. For a literature review on early works, interested readers can refer to [1]. Among the recent developments, level set segmentation and graph-based segmentation are the two popular methods that have been received a number of attentions. The level set method is basically a reformulation of the active contour model [2] in the framework of level set. In spite of the increase in dimension, this formulation does not require a parameterized representation of the tracking object in the course of curve evolution. More importantly, topology change is very easy to adapt the contour towards the boundary of objects with varying shapes. These salient features make the method particularly useful in tracking interfaces and shapes [3]. Different from most active contour models that relate the object boundaries to image gradients, Chan and Vese [4] proposed a model where the stopping term is related to a particular segmentation of the image. Another advantage of this formulation is that the initial curve can be anywhere in the image. In most of traditional level set methods, it is usually require a step to periodically re-initialize the level set function to a signed distance function throughout the process of evolution in order to maintain the evolution stable. In [5], a new variational formulation was presented to force the level set function to be close to a signed distance function without the step of re-initialization. Another type of segmentation methods that has been of intensive interest in recent years is the graph-based method, in particular due to the efficient algorithm by Boykov and Kolmogorov [6] for computing the max-flow for computer vision related graph. The method takes image pixels as graph nodes and lines linking pairs of pixels as graph edges, thus the segmentation problem can be represented in terms of a graph. Wu and Leahy introduced a minimum graph cut method for image segmentation, but the method tends to bias towards finding small components. Shi and Malik [7] proposed a normalized cuts method to address this bias issue. In spite of the good performance as reported, the method yields an NP-hard computational problem and is pretty time consuming for real-time applications. A method running in ( ) m m log Ο for m graph edges was presented in [8], where the segmentation is based on pairwise region comparison with decision following the global properties of being not too coarse and not too find according to a particular region comparison function. Despite the tremendous efforts in object detection, it remains a challenging issue for reliably detecting objects under a wide range of variation in scene view, in particular for those computer vision systems running in outdoor and uncontrolled environment. A typical difficulty is the insufficient dynamical range for some observed objects, which would make the object appear with low contrast and barely visible even with human observation. The insufficient dynamical range could be due to several reasons. Very often it results from imperfect lighting condition, such as directional strong lighting, non-uniform weak lighting. Poor weather (like fog) or the medium (like turbulent water or some solvent) where the object is immersed could lead to the rapid decrease on the lighting transmission from objects to image plane on the one hand, and more importantly the effect of strong lighting interference due to particle reflection on the other hand. As a result, objects with different depth could exhibit remarkably different contrasts. For these images, normally it would require a good enhancement before proceeding to the step of object detection, which is another nontrivial issue. Here we present a simple yet MVA2011 IAPR Conference on Machine Vision Applications, June 13-15, 2011, Nara, JAPAN 4-10