Multiresolution visual enhancement of hazy underwater scene

Abstract

Haze is an obvious phenomenon in the underwater scenario. The scene visibility reduces to a great extent due to haze, which makes the underwater visual surveillance quite a challenging task. In this article, we have exploited the multi-resolution ability of discrete wavelet transform and applied dark channel prior based transmission map estimation scheme to dehaze the highly degraded underwater image and restored the color. A three-fold scheme for dehazing of underwater sequences is proposed. In the first stage, image details are extracted using discrete wavelet transform followed by image negative operation. In the second stage, the negative of detail images are enhanced by the help of dark channel prior. The third stage is used for reconstruction, where the enhanced image details are used along with the single level approximate of the input image to get the dehazed underwater image using inverse discrete wavelet transform. The proposed scheme is tested with numerous standard underwater images, as well as the excavation images of Dwaraka (Dvārakā) underwater ruins. The effectiveness of the proposed scheme is justified by comparing it with different state-of-the-art image dehazing techniques. The quantitative evaluation has been carried out using five well established general purpose non-reference image quality indices namely BIQI (blind image quality index), BLIINDS (BLind Image Integrity Notator using DCT Statistics), DIIVINE (Distortion Identification-based Image Verity and INtegrity Evaluation), BRISQUE (Blind/Referenceless Image Spatial Quality Evaluator), and SSEQ (Spatial-Spectral Entropy-based Quality). Encouraging scores of 35.01, 30.105, 27.22, 30.10, and 27.8, are achieved for the BIQI, BRISQUE, SSEQ, DIIVINE, and BLIINDS, respectively. Four evaluation measures, exclusively designed for underwater scenarios (underwater image quality, contrast, sharpness, and colorfulness measures) are also used to test the performance of the proposed scheme.

Appendix: Evaluation Measures

Quantitative analysis of the images is carried out using the non-reference image quality assessment measures. The reason for choosing these measures is that, in our experiments, the reference or ground truth images are not available.

BRISQUE :

[38]: It is a natural scene statistic based distortion-generic blind/no-reference image quality assessment model which operates in the spatial domain. It uses scene statistics of locally normalized luminance coefficients to quantify possible losses of ‘naturalness’ in the images due to the presence of distortions, thereby leading to a holistic measure of quality. A support vector regression (SVR) model is employed by using the differential mean opinion score (DMOS) value. The measure is based on the mean subtracted contrast normalized (MSCN) coefficient, which is given by;

$$ I_{MSCN} = \frac{I(i,j)-\mu(i,j)}{\sigma(i,j)+c}, $$

(24)

where,

$$ \mu(i,j)={\sum}_{k=-K}^{K}{\sum}_{l=-L}^{L}w_{k,l}I(i+k,j+l), $$

(25)

$$ \sigma(i,j)=\sqrt{{\sum}_{k=-K}^{K}{\sum}_{l=-L}^{L}w_{k,l}\left[I(i+k,j+l)-\mu(i,j)\right]^{2}}. $$

(26)

DIIVINE :

[40]: It is a distortion-agnostic approach to blind image quality assessment that utilizes concepts from natural scene statistics to not only quantify the distortion but also quantify the distortion type afflicting the images. First, it computes the wavelet coefficients and normilize them. Next, it identifies the distortion afflicting the images and then performs the distortion-specific quality assessment.

BLIINDS :

[52]: It is an efficient, general-purpose, non-distortion specific, blind/no-reference image quality assessment algorithm that uses discrete cosine transform coefficients to perform the distortion-agnostic quality assessment.

BIQI :

[39]: It is a new two-step framework for no-reference image quality assessment based on natural scene statistics. Once trained, the framework does not require any knowledge of the distorting process and the framework is modular so it can be extended to any number of distortions.

SSEQ :

[35]: It is an efficient general-purpose no-reference image quality assessment model that utilizes local spatial and spectral entropy features on distorted images. It Uses a 2-stage framework of distortion classification followed by a quality assessment. It utilizes a support vector machine to train an image distortion and quality prediction engine. It is capable of assessing the quality of a distorted image across multiple distortion categories.

UICM :

[47]: This parameter is devised to measure the colorfulness of the image. This is given as;

$$ UICM = -0.0268 \sqrt{\mu_{\alpha, RG}^{2}+\mu_{\alpha, YB}^{2}}+0.1586 \sqrt{\sigma_{\alpha, RG}^{2}+\sigma_{\alpha, YB}^{2}}, $$

(27)

where μ represents the chrominance intensity, and σ² represents the variance in the chrominance.

UISM :

[47]: The underwater image sharpness measure aims to compute the measure of sharpness of an image. It is computed by;

$$ UISM = {\sum}_{c=1}^{3}\lambda_{c} EME \left( GrayscaleEdge_{c}\right), $$

(28)

$$ EME = \frac{2}{k_{1}k_{2}}{\sum}_{l=1}^{k_{1}}{\sum}_{k=1}^{k_{2}}log\left( \frac{I_{max,k,l}}{I_{min,k,l}}\right), $$

(29)

where k₁k₂ is the number of blocks an image got divided by.

UIConM :

[47]: It is the underwater image contrast measure, which is defined mathematically as;

$$ UIConM = logAMEE(Intensity), $$

(30)

where the AMEE is the Agaian measure of enhancement by engropy. This is the average michaelson contrast in the local regions.

UIQM :

[47]: It is the normalized underwater image quality measure. It is given by:

$$ UIQM = c_{1} \times UICM + c_{2} \times UISM + c_{3} \times UIConM, $$

(31)

where c₁, c₂, and c₃ are the contributing factors which are set to 0.0282, 0.2953, and 3.5753, respectively.