Visual enhancement of underwater images using Empirical Mode Decomposition

Abstract

Most underwater vehicles are nowadays equipped with vision sensors. However, it is very likely that underwater images captured using optic cameras have poor visual quality due to lighting conditions in real-life applications. In such cases it is useful to apply image enhancement methods to increase visual quality of the images as well as enhance interpretability and visibility. In this paper, an Empirical Mode Decomposition (EMD) based underwater image enhancement algorithm is presented for this purpose. In the proposed approach, initially each spectral component of an underwater image is decomposed into Intrinsic Mode Functions (IMFs) using EMD. Then the enhanced image is constructed by combining the IMFs of spectral channels with different weights in order to obtain an enhanced image with increased visual quality. The weight estimation process is carried out automatically using a genetic algorithm that computes the weights of IMFs so as to optimize the sum of the entropy and average gradient of the reconstructed image. It is shown that the proposed approach provides superior results compared to conventional methods such as contrast stretching and histogram equalizing.

Introduction

In recent years, there is an increasing interest in underwater remotely operated vehicles (ROV’s) and autonomous underwater vehicles (AUV) for submarine and military operations. These vehicles are typically equipped with optical sensors (cameras) to capture underwater images for short range applications. The main drawback of employing optical cameras in underwater applications is limited visibility that can be restricted to about twenty meters in clear water and less than a few meters in turbid and coastal waters (Arnold-Bos, Malkasse, & Kervern, 2005).

The quality of images acquired under the sea can be poor because of specific propagation properties of light in water, so image enhancement is necessary to enable effective interpretation means for operators. The most important reason of underwater image degradation is due to transmission properties of light in water, such as absorption (light disappears) and scattering (light changes direction). Another problem is related to depth. Due to the nature of underwater optics, red color disappears at the depth of about 3m and orange color starts diminishing a little further, while yellow color is lost at a depth of about 10 m and green goes off at further depths, so finally at 25 m only blue color remains, thus causing the image to typically appear bluish (Abril, Méndez, & Dudek 2005). The third problem is marine snow that resembles snowflakes suspended in the deep ocean. Marine snow can produce bright artifacts. As a result, underwater images have restricted visibility, non-uniform lighting, low-contrast, diminished colors and blurring of image features. Some of these effects are even observed when external lighting is used. Enhancement methods have been proposed in the literature to improve image quality, compensate attenuation effects, enhance contrast, adjust colors, suppress noise and blur, while preserving and possibly even enhancing edges.

In the literature, the number of methods proposed for or applied to underwater images for enhancement and pre-processing is rather limited. The most widely used enhancement methods are conventional histogram equalization (Thakur & Tripathi, 2010) and contrast stretching (Iqbal, Abdul Salam, Osman, & Talib, 2007) approaches. (Iqbal et al., 2007) presented an underwater image enhancement method using an integrated color model. Alternative methods proposed in the literature are adaptive smoothing techniques, and some filtering methods such as homomorphic filtering, anisotropic filtering, and wavelet denoising (Bazeille et al., 2006, Padmavathi et al., 2010).

In this paper, it is proposed to utilize an Empirical Mode Decomposition (EMD) based enhancement method for underwater images. EMD is a signal decomposition technique which is particularly suitable for the analysis of non-stationery and non-linear data (Huang et al., 1998). In EMD, the signal is decomposed into components called Intrinsic Mode Functions (IMFs) and a residue. The lower order IMFs capture fast oscillation modes (high spatial frequencies in images) while higher order IMFs typically represent slow spatial oscillation modes (low spatial frequencies in images). EMD has some important advantages compared to Wavelet (Janusauskas, Jurkonis, Lukosevicius, Kurapkiene, & Paunksnis, 2005) and Fourier transform (Zhidong & Yang, 2007) techniques. Although many real-life systems are non-linear and non-stationery, the data is assumed to be stationary and linear in the Fourier Transform. In the wavelet transform it is possible to use different wavelet types and the performance can change according to this selection. On the other hand, EMD does not have basis functions and decomposes the signal based on its intrinsic properties. IMFs can contain both high and low frequency details at different signal locations depending on the signal characteristics.

EMD has recently started to find a wide area of utilization in signal processing applications. EMD is for example applied to signals such as EEG (Weng, Blanco-Velasco, & Barner, 2006), cardiotocograph (Krupa, Mohd Ali, & Zahedi, 2009) in biomedical signal processing. In Janusauskas et al. (2005) EMD and wavelet decomposition are used to detect human cataract using ultrasound signals. EMD is applied to 2D face images as a pre-processing step to remove illumination artifacts for a face recognition application in (Bhagavatula & Savvides, 2007). EMD is used for image compression in (Linderhed, 2004). EMD has been successfully applied to hyperspectral images in Demir & Ertürk (2008). Liu, Liao, and Sang (2005) presents an algorithm for removing noise in sonar images using EMD. It has been proposed to use EMD for increased target detection capabilities in sonar images in Taşyapı Çelebi and Ertürk (2010a).

In Hariharan, Gribok, Abidi, and Koschan (2006), EMD has been utilized for image fusion and enhancement, where images from different imaging modalities are decomposed into their IMFs and fusion is performed at the decomposition level and IMFs are fused with different weights to obtain the reconstructed image. Not only it has been adopted this approach for underwater images in Taşyapı Çelebi and Ertürk (2010b). In this paper, a novel genetic algorithm based procedure is presented that obtains IMF weights for automatic processing.