Histogram equalization and optimal profile compression based approach for colour image enhancement

https://doi.org/10.1016/j.jvcir.2016.04.019Get rights and content

Highlights

  • A pipeline approach to increase contrast and restore vividness to colour images.

  • Pre-processing procedures include colour stretching, and histogram equalization.

  • A magnitude compression and a saturation maximization stage as post processing.

  • Image content based feedback provides optimal compression to reduce artefacts.

  • Comprehensive assessment shows improvements on contrast and colour vividness.

Abstract

Many vision based applications depend on images with sufficiently high contrast and colourfulness so that ample amount of information is available to accurately describe objects captured in an image scene. Poor image capturing conditions are often unavoidable but can be compensated. Approaches based on intensity histogram equalization are popular to increase the information content within an image but over-enhancement often results in the production of unwanted artefacts. Furthermore, when constrained to only an intensity-based enhancement, insufficient enrichment on colourfulness and saturation is often observed. In order to address these limitations concurrently, a pipelined approach that incorporates a colour channel stretching process, a histogram equalization step, a magnitude compression procedure, and a saturation maximization stage is proposed. Quantitative and qualitative results obtained from experiments on a wide variety of natural scene images demonstrate the effectiveness of the proposed approach over other methods at reducing artefact while increasing image contrast and colourfulness.

Introduction

Digital colour images are important medium to capture information of a scene observed by a viewer. Once captured, vision based processing methods are used to manipulate or enhance images for use in a variety of science and engineering disciplines. For instance, contrast enhancement and image denoising are important in medical diagnosis through images [1]. The remote sensing [2] field of study depends on images and enhancement procedures to aid activities such as resource exploration and land cover surveillance. In robotics, digital cameras are frequently used to complement various sensors [3] for navigation and control. For the monitoring of machine operation conditions, image based ferrography has found many applications in practice [4]. While high performing results are expected from image based technologies, the enhancement of image quality is a prevalent priori requirement. In addition to aiming at the increment in image contrast, it is inextricable that colour quality should also be considered.

The enhancement of colour image quality can be carried out by transforming the image from the primary colour space, namely red, green and blue, to the human perceptible domain of hue, saturation and intensity. In this class of methods, hue is mostly retained for faithful reproduction of the scene information. Increments in saturation improve the vividness of the objects captured in the image. Moreover, in most cases, the intensity is manipulated to provide a boosted contrast that aids the extraction of information conveyed from the scene. In this context, the histogram equalization technique is an attractive candidate for its implementation simplicity [5]. However, the development of a generic equalization algorithm still remains challenging due to the fact that different images contain varying amount of information contents.

With regard to histogram equalization, it can be viewed as a statistical re-allocation of intensities such that all permitted levels are used to mimic scene information in the image. Practically, this is realized in the form of histogram modifications [6]. Variations in this class of methods include incorporating local features around a neighbourhood for each pixel to aid equalization [7] or through the matching of an input histogram to a target histogram. The target histogram is selected in accordance with some desirable objectives such as increasing the image sharpness [8]. Another alternative means to specify the target histogram was reported in [9], whereby histogram smoothing was performed before matching.

Manipulations on the histogram are often adopted for enhancing the image contrast. There had been different approaches proposed for different application areas. In [10], thresholds were imposed to limit the intensity change before and after equalization on infrared images. The improvement on image perception was obtained from dynamic adjustments using the relationship between a centre and surrounding pixels [11]. The principle of sectoring histograms was further modified for contrast enhancement with clippings on the histogram peaks [12].

The usage of brightness as an enhancement objective had received a lot of recent attention. A sub-histogram based approach was developed and reported in [13] where the adverse effects of under and over exposure were reduced. A similar approach was also reported in [14] where a recursive loop was included. In [15], the authors used the selection of multiple histogram references to achieve their brightness objective. Another approach is based on the division of histograms, whereby narrow segments were extended to the full dynamic range, resulting in the increase of contrast [16]. Other researchers tried to balance brightness on two separated sections derived from the original histogram [17]. There is no universal rule proposed so far in dividing the histogram. In [18], the authors divided the histogram into three sectors using the modified Otsu threshold before applying equalization. A variation was proposed in [19] where three histogram segments were obtained and then smoothed using a Gaussian function to form the reference for equalization. In [20], the division was based on the quartiles and modified by a power law. The resultant histogram was used as the target in the equalization process. In [21], the division of histograms were not confined to user specifications but was made on the basis of the number of peaks. Other than hard dividing the histogram, a Gaussian mixture model was employed in [22] to represent the original histogram. Then the Gaussians were shifted and scaled for use in equalization. An integrated approach was presented in [23] for contrast enhancement as well as brightness preservation. The work reported in [24], on the other hand, made use of the pixel gradient to create a reference histogram instead of directly using the intensity. Although satisfactory results were obtained from the aforementioned enhancement methods by modifying the histograms, there are still cases that histogram based approaches would produce unwanted artefacts. These include the occurrence of un-naturalness and loss of colour saturation.

In order to enhance the quality of colour images, encompassing contrast, sharpness and colourfulness, a pipelined approach is proposed in this work. The given colour image is first stretched in each of its primary colour channels providing an enriched colour vividness. The colour image is further converted from the primary colours to de-coupled human perceptual domains. Next, the intensity is manipulated using an improved form of histogram equalization whereby unwanted artefacts are compressed using an optimally chosen hyperbolic tangent profile. Instead of directly outputting the equalized image, the colour saturation is restored and maximized in accordance to the input image to mitigate the loss of colourfulness.

The rest of the paper is organized as follows. In Section 2, the proposed pipelined approach is presented in the system architecture, along with detailed description of the features within each block, starting from the input image on the left to the output image on the right. Results from an experiment conducted is detailed in Section 3. This is followed by a discussion in Section 4. Finally, the conclusion drawn is given in Section 5.

Section snippets

Proposed approach

Consider an input colour image I(u,v)={R(u,v),G(u,v),B(u,v)}, where {R,G,B} are the red, green and blue colour channels, (u,v) are the pixel coordinates, u=1,,U,v=1,,V, and U,V are the width and height of the image. The total number of pixels in the image is N=U×V. A system block diagram of the proposed pipelined approach is depicted in Fig. 1. The operations carried out in each individual functional block are described in the following subsections.

Experiments

Experiments were conducted to verify the performance of the proposed equalization compression method. A collection of 200 colour images of natural scenes were used. The size of each image is 400×300 width-by-height. The images were stored in 24-bit colour JPG format which were read and processed in Matlab. Based on several pilot tests, the profile gain factor was initially set to [1,5] before being optimized. In order to fairly evaluate the performance of our approach, notable histogram based

Discussion

As the gradient and colourfulness values of local enhancement techniques (Adp Eq and Smooth Eq) are higher than our proposed local enhancement method (Prop Eq), a further discussion is required to show that our pipeline approach is effective at maximizing information content without distorting local colours and features. In a separate experiment, the uniform histogram equalization, adaptive contrast equalization, and smooth histogram equalization methods were enhanced with pre-processing

Conclusion

A framework to improve the quality of colour images was developed and presented in this paper. Through a series of experiments using qualitative assessment and quantitative measures such as entropy, gradient, colourfulness and saturation, it was highlighted that the proposed pipeline approach increases image content, improves contrast, and sharpens colour images without compromising on colourfulness and saturation. Unwanted artefacts that commonly features during histogram equalization were

References (28)

  • G. Zhang et al.

    Super pixel based remote sensing image classification with histogram descriptors on spectral and spatial data

  • F. Su et al.

    Adaptive colour feature identification in image for object tracking

    Math. Probl. Eng.

    (2012)
  • J.A. Stark

    Adaptive image contrast enhancement using generalizations of histogram equalization

    IEEE Trans. Image Process.

    (2000)
  • S. Poddar et al.

    Non-parametric modified histogram equalisation for contrast enhancement

    Image Process. IET

    (2013)
  • Cited by (0)

    This paper has been recommended for acceptance by Zicheng Liu.

    View full text