A histogram specification technique for dark image enhancement using a local transformation method
- 1.4k Downloads
Traditional image enhancement techniques produce different types of noise such as unnatural effects, over-enhancement, and artifacts, and these drawbacks become more prominent in enhancing dark images. To overcome these drawbacks, we propose a dark image enhancement technique where local transformation of the pixels have been performed. Here, we apply a transformation method of different parts of the histogram of an input image to get a desired histogram. Afterwards, histogram specification technique has been done on the input image using this transformed histogram. The performance of the proposed technique has been evaluated in both qualitative and quantitative manner, which shows that the proposed method improves the quality of the image with minimal unexpected artifacts as compared to the other techniques.
KeywordsDark image enhancement Histogram equalization Histogram specification
Histogram equalization (HE) is a simple and effective contrast enhancement technique for enhancing an image. HE spreads the intensities of an image pixels based on the whole image information. As a result, there might be a case where some low occurring intensities are transformed to become merged with neighboring high occurring intensities, which creates over-enhancement [2, 3]. Moreover, mean shift problems may also occur in such cases. So, brightness preservation cannot be guaranteed in HE. An improved version of HE is the brightness preserving bi-histogram equalization (BBHE)  which produces better results as compared to HE, and most often, it preserves the brightness of the original image. However, it may not give desired outcome when the image pixel distribution does not follow the symmetric distribution . A method called equal area dualistic sub-image histogram equalization (DSIHE)  performs better than BBHE because it separates the histogram based on the image median. Recently, Rahman et al. proposed an enhancement technique where the histogram is divided using harmonic mean of the image, and then, HE has been applied . Still, these techniques may not always increase image contrast, especially for dark images, since the focuses of these methods are the preservation of the brightness.
A combination of BBHE and DSIHE is recursively separated and weighted histogram equalization (RSWHE) , and it preserves the brightness and enhances the contrast of an image. The core idea of this algorithm is breaking down a histogram into two or more portions and then applying a weighting function on each of the sub-histograms, on which the histogram equalization is performed. However, some statistical information might be lost after the histogram transformation, and the desired enhancement may not be achieved . Inspired by the RSWHE, the adaptive gamma correction with weighting distribution (AGCWD)  uses gamma correction and luminance pixel probability distribution to enhance the brightness and preserves the available histogram information. Here, a hybrid histogram modification (HM) technique is used to combine the traditional gamma correction (TGC) and traditional histogram equalization (THE) methods. Although this method enhances the brightness of the input image in most of the cases, it may not give satisfactory results when an input image has lack of bright pixels.
An extended version of BBHE and DSIHE is the minimum mean brightness bi-histogram equalization (MMBEBHE)  where input image histogram is recursively separated into multiple sub-histograms using absolute mean brightness error (AMBE). Although this technique performs good in contrast enhancement, MMBEBHE incurs much side effects. ChaoWang and Zhongfu Ye  propose the brightness preserving histogram equalization with maximum entropy (BPHEME), which provides acceptable results for continuous case, but fails for discrete ones. Chao Zuo et al.  propose range limited bi-histogram equalization (RLBHE) that preserves the mean brightness of the image. However, the computational complexity is very high as compared to the other methods. Huang et al. address this problem by proposing a hardware-oriented implementation . Bilateral Bezier curve (BBC) method works in dark and bright regions of an image separately .
The Histogram Modification Framework (HMF)  focuses on minimizing a cost function to get the target histogram. SM Pizer et al. propose weighted adaptive histogram equalization (WAHE) , which processes the input image based on the impact of pixels to the histogram by considering the closeness of the pixels. Although WAHE improves image contrast, it requires intensive computation. Content-aware channel division (CACD) is proposed in  for dark images enhancement. CACD groups the image information with common characteristics. However, only grouping the contrast pairs into intensity channels may not always be sufficient because different intensity channels may possess same characteristics .
Some renowned histogram specification-based methods such as automatic exact histogram specification (AEHS)  and dynamic histogram specification (DHS)  are available for image enhancement. AEHS is proposed for both local and global contrast enhancement whereas DHS uses differential information from an input histogram to eliminate the annoying side effects. Besides the aforementioned enhancement methods, there exist few other methods such as guided image contrast enhancement  and image enhanccement by entropy maximization , contrast enhancement based on piece-wise linear transformation (PLT) , layered difference representation (LDR) , and inter pixel contextual information (named as CVC) . LDR first divides the gray levels of an image into different layers and makes a tree structure for deriving a transformation function. After getting the transformation functions for each layer, all of those are aggregated to achieve the final desired transformation function handling the sudden peaks. However, it may not perform accurately in some cases of dark images.
Few methods are solely developed for dark image enhancement, but their performances are not always satisfactorily. The main limitation is that these techniques may transfer most of the pixels from dark region to the bright region which might cause over-enhancement and unwanted shift in brightness. To mitigate this, we propose a method specifically to enhance images having dark portions in them (a preliminary version of this work can be found in ). We divide the whole image histogram into several segments and then modify those to have a histogram with desired characteristics. Finally, histogram specification is performed on the input image using this desired histogram. According to our method, the desired histogram is carefully created to enhance the image, especially the dark parts. Moreover, the gray levels of one segment are not transferred to another segment. This helps to avoid over-enhancement and additional noise. Experimental results also advocate for the effectiveness of the proposed method.
Rest of the paper is organized as follows. Section 2 discusses the proposed method, and the results are presented in Section 3 using both qualitative and quantitative analysis. Finally, Section 4 concludes the contribution of this paper.
2 Proposed method
In HSV color model, V (luminance) and color information (hue and saturation) are decoupled.
Color relationships of HSV color model are described more meticulously than RGB color model.
We can easily transform RGB color model to HSV color model and vice versa.
After applying the enhancement on V channel, the image is converted back to RGB. The whole enhancement process consists of two major steps: (1) preprocessing and (2) enhancement. These two are discussed in the following sub-sections.
Here, σ represents standard deviation. In an image, specifically in a dark image, some intensities might exist at very few number of pixels that are not important to visualize the objects in that image. Hence, we propose to merge the histogram bins of such unimportant intensities with the neighboring bins so that our later procedures get more room to process the histogram segments.
We use Algorithm 1 to dynamically calculate a threshold (τ) from the input histogram and to find such insignificant intensities. τ gives a level for the accumulation of a bin justifying whether the presence of the corresponding intensity is significant or not. Algorithm 2 scans every bin of the histogram and compares it with τ. If the accumulation is less than τ, the accumulation is added to the next bin’s accumulation. Thus, we dissolve the insignificant bins in a histogram.
From the preprocessing step, we obtain prominent peaks and valleys for the histogram of a given image. These peaks and valleys are used to identify a set of segments where a segment is defined as follows:
For enhancement, we perform two types of operations for each of these segments of the histogram: First, segment reallocation and second, gray level transformation within segment. Detail of these two steps are described in the following sub-sections.
2.2.1 Segment reallocation
where N = total number of segments, D(i) = shifting distance of ith segment, x i = accumulation of ith segment.
For example, if the total number of accumulation =100, accumulation of the ith segment =10 and total number of segments =50, then the shifting distance will be D(i)=(10/100)×50=5. So, new ending position of this segment will be Vi+1+5. And thus, the width of a segment depends on the number of pixels it contains. Here, the segments containing more pixels will be allocated more gray levels, which are expected for enhancement.
where l i = width of the ith segment, L i = resultant width of the ith segment, θ = ending bin position of the last segment. Thus, it also gives more space to perform intra-segment transformation.
2.2.2 Intra-segment transformation
where T(i) = transformed intensity of ith bin, Ω(i) = accumulation of ith bin, s = starting bin position of the segment, k = last bin position of the segment, L = width of the corresponding segment (from Eq. 4).
2.2.3 Histogram specification
By performing the segment reallocation and intra-segment transformation, we get the the transformed histograms for both edge and non-edge images. The desired histogram is obtained by combining these two histograms which is used to perform histogram specification.
We apply Eq. 7 on every gray level of the input image to get the enhanced image. The output image is transformed back to RGB if the original image is RGB.
3 Experimental results and discussion
In this section, the results of the proposed technique has been compared to the existing state-of-the-art image enhancement techniques, namely HE , AEHS , CVC , LDR , WAHE , CACD , AGCWD , and RSWHE . The comparison has been performed in both qualitative and quantitative manner. To evaluate the proposed method, images are taken from CACD , Caltech  and UIUC Sport Event . The details are presented in the following sections.
3.1 Qualitative measures
In “girl” image, the main challenge is to increase overall brightness without incurring artifacts on hair and necklace. Most of the methods cannot reveal the detail texture and girl’s necklace. HE and AEHS over-enhance the image and produce lots of artifacts. CACD produces comparatively good result, but the image is not properly illuminated. LDR and WAHE cannot increase the brightness properly, and the original color of the image is also degraded whereas our proposed “girl” image increases brightness and preserves original color contrast. Thus, our proposed method performs better than the others in this respect.
In the “fountain” image, the challenges are to preserve the lamp as original as possible and keep naturalness of the grass. Here, HE and AEHS over-enhance the wall and grass. Glasses of windows do not look original. LDR, CVC, WAHE, and CACD cannot enhance the image properly and the output images are still a little bit dark. On the other hand, HE and AEHS increase brightness at a large rate which incurs artifacts on the wall. However, the proposed method preserves the brightness of lamp and the natural color of grass. Thus, the proposed method produces comparatively better result than others.
The desired enhancement of “street1” image is to increase the brightness in such a way that hidden information of the image can be extracted. CVC, LDR, WAHE, CACD, AGCWD, and RSWHE cannot enhance the brightness of this image properly because the hidden information of “three man” in the image is unclear. However, the proposed method increases the brightness, and hidden information is more clearly visible than the others.
The challenges of the “dark ocean” image are to preserve the ray of sunlight and enhance image contrast properly. HE and AEHS increase image brightness, but naturalness of sunshine is degraded. LDR and CVC cannot enhance the brightness of this image properly. The resultant images of these two methods are still significantly dark. CACD and AGCWD preserve the sunshine but loss the naturalness of the water whereas RSWHE degrades the quality of sun rays. In this case, the proposed approach preserves the sunshine and increases overall image brightness.
3.2 Quantitative measures
For the purpose of quantitative evaluation, fifty images are taken from aforementioned three datasets namely CACD, Caltech and UIUC Sport Event. To evaluate these images quantitatively, we use Root mean square (RMS), structural similarity, and perceptual quality metric (PQM) metrics which are discussed in the following sub-sections.
3.2.1 RMS contrast
Here, M and N are the image dimension. Ii,j, μ represent pixel intensity and mean of an image. Usually, larger value of the RMS represents better quality of the image. However, for enhancing a dark image, this might not be true because the increase of RMS depends on the increase of the diversified intensity which may also deteriorate the image quality with increased number of artifacts. This is also observed when we measure the average values of the outputs of different methods for 50 dark images. The average RMS values obtained by AGCWD, CVC, and LDR are 0.29, 0.27, and 0.26, respectively. On the other hand, the proposed method and CACD obtain only 0.25 and 0.27, respectively, though their outputs are qualitatively more soothing and better in terms of other measures such as PQM. Hence, the RMS values does not actually reflect the enhancement for dark images.
where α, β, γ1, γ2, and γ3 are the model parameters. A, B, and Z are the features (for details, please see ). The average PQM calculated for the 50 enhanced images is the highest (9.03) for the proposed method. CACD also shows a very competitive PQM (8.94). The lowest PQM is found for HE (8.73), and it is 8.82 for AGCWD. According to these results, it is also clear that the proposed method produces better output as compared to the other image enhancement techniques in most of the cases.
In this paper, a locally transformed histogram-based technique has been proposed for dark image enhancement. This method works better as compared to other methods because we do not apply our transformation method on the whole histogram of an input image. Rather, our transformation method is applied on a small segment of the input image histogram. As a result, the proposed technique does not get affected from over-enhancement problem. Our experimental results show that it reaches higher performance metrics as compared to existing techniques.
This work is supported by the University Grants Commission, Bangladesh under the Dhaka University Teachers Research Grant No Reg/Admin-3/54290.
All the authors have contributed in designing, developing, and analyzing the methodology, performing the experimentation, and writing and modifying the manuscript. All the authors have read and approved the final manuscript.
The authors declare that they have no competing interests.
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
- 2.Gonzalez RC, Woods RE (2002) Digital image processing. Prentice Hall, Upper Saddle River.Google Scholar
- 5.Rahman S, Rahman MM, Hussain K, Khaled SM, Shoyaib M (2014) Image enhancement in spatial domain: a comprehensive study In: Computer and Information Technology (ICCIT), 2014 17th International Conference On. IEEE. pp 368–373 Google Scholar
- 7.Amil FM, Rahman MM, Rahman S, Dey EK, Shoyaib M (2016) Bilateral histogram equalization with pre-processing for contrast enhancement In: Software Engineering, Artificial Intelligence, Networking and Distributed Computing (SNPD), 17th IEEE International Conference on. IEEE, ShanghaiGoogle Scholar
- 18.Priyakanth R, Malladi S, Abburi R (2013) Dark image enhancement through intensity channel division and region channels using Savitzky-Golay filter. Intl J Sci Res Publications ISSN 3(8):2050–2016.Google Scholar
- 24.Lee C, Lee C, Kim CS (2012) Contrast enhancement based on layered difference representation In: Image Processing (ICIP), 19th IEEE International Conference On. IEEE, Florida. pp 965–968Google Scholar
- 26.Hussain K, Rahman S, Khaled SM, Abdullah-Al-Wadud M, Shoyaib M (2014) Dark image enhancement by locally transformed histogram In: Software, Knowledge, Information Management and Applications (SKIMA), 8th International Conference On. IEEE, Dhaka. pp 1–7Google Scholar
- 28.Griffin G, Holub A, Perona P (2007) Caltech-256 object category dataset. Technical Report CNS-TR-2007-001, California Institute of Technology. http://authors.library.caltech.edu/7694/.
- 29.Li L-J, Fei-Fei L (2007) What, where and who? Classifying events by scene and object recognition In: Computer Vision, 2007. ICCV 2007. IEEE 11th International Conference On. IEEE, Rio de Janeiro. pp 1–8Google Scholar
- 32.Wang Z, Sheikh HR, Bovik AC (2002) No-reference perceptual quality assessment of JPEG compressed images In: Image Processing. 2002. Proceedings. 2002 International Conference On, vol. 1. IEEE, New York. p 477Google Scholar
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.