Image contrast enhancement via entropy production

Abstract

In this paper a novel approach for image contrast enhancement is proposed. Employing a thermodynamical model to define local information content, the method derives a measure of local contrast which takes into account spatial structure across multiple scales. From this measure a local contrast map is computed, which is then applied to the image, giving rise to a selective enhancement of the original image. Here some applications to medical images will be presented as well as a comparison with other methods.

Introduction

Generation of digital images involves several steps like formation, digitization, transmission, which may affect the information content of the image. One of the most common degradations in the recorded image is its poor contrast—roughly, the difference between its highest and lowest intensity values—due for example, to a poor illumination of the object or the scene to be imaged [1]. In particular, digital image enhancement techniques have been widely used in medical imaging in order to enhance contrast of image features. Also, it is worth noting that an enhancement technique performing well in enhancing biomedical images may not be as good in enhancing satellite images.

Contrast enhancement methods can be broadly classified into global, local and multiscale. Global methods basically rely on the modification of the histogram such as in histogram equalization and histogram specification [1]. Clearly such methods suffer from the drawback that the image characteristics differ considerably from one region to another in the same image and the local histogram may not follow the global histogram. This indeed occurs when complex images such as medical images are considered; in this case, it is reasonable to resort to context-sensitive techniques relying on local contrast variation. In local methods of contrast enhancement, a definition of local contrast is used to measure the contrast and enhance the image by modifying the contrast measurement [2], [3], [4]. In particular, Beghdadi and Le Negrate [5] first defined the contrast by considering edge detection operators. The main drawback of local approaches is that they are single-scale spatial domain methods and thus can only enhance the contrast of a narrow range of sizes, as determined by the size of the local processing region. In order to enhance features of all sizes simultaneously, multiresolution enhancement methods have been proposed [6], [7] based on the wavelet transform [8]. As an alternative to the wavelet representation, a multiscale morphological approach to local contrast enhancement has been recently developed [9], while Boccignone et al. [10], [11] have suggested a multiscale approach using a nonlinear scale-space representation of the image generated by anisotropic diffusion [12].

This paper contributes a novel approach. The idea is to perform contrast enhancement relying on the structural information contained within a given region of an image, which can be encoded by considering the dynamical loss of information. The latter is measured by the production of entropy, in the thermodynamical sense, during a fine-to-coarse transformation, namely a nonlinear scale-space evolution of the original image. [13]. The evolution of entropy production along the transformation, locally characterizes different kinds of features such as edges and texture, and is eventually used to form a contrast map, which is the basis of the proposed enhancement method.

The outline of this paper is as follows. In Section 2, we provide a brief overview of key concepts related to contrast enhancement. In Section 3 we discuss the theoretical background of the approach, in particular its links to scale-space theory. In Section 4, the enhancement method is described providing a detailed discussion as regards the behavior of contrast in anisotropic scale-space. In Section 5 we present the experimental work. We supply examples obtained within the medical imaging field, showing that the proposed method encompasses advantages peculiar to locally adaptive and other multiscale techniques, while allowing to capture the nonlinear effects of contrast sensitivity. We draw some conclusions in Section 6.