Saliency motivated pulse coupled neural network for underwater laser image segmentation

Abstract

The detection range of underwater laser imaging technology achieves 4-6 times of detection range of conventional camera in intervening water medium, which makes it very promising in oceanic research, deep sea exploration and robotic works. However, the special features in underwater laser images, such as speckle noise and non-uniform illumination, bring great difficulty for image segmentation. In this paper, a novel saliency motivated pulse coupled neural network (SM-PCNN) is proposed for underwater laser image segmentation. The pixel saliency is used as external stimulus of neurons. For improvement of convergence speed to optimal segmentation, a gradient descent method based on maximum two-dimensional Renyi entropy criterion is utilized to determine the dynamic threshold. On the basis of region contrast in each iteration step, the real object regions are effectively distinguished, and the robustness against speckle noise and non-uniform illumination is improved by region selection. The proposed method is compared with four other state-of-the-art methods which are watershed, fuzzy C-means, meanshift and normalized cut methods. Experimental results demonstrate the superiority of our proposed method to allow more accurate segmentation and higher robustness. Key words: underwater laser image, pulse coupled neural network, pixel saliency, region contrast CLC number: TP 183 Document code: A IntroductionThere are quite limited means of detecting objects in underwater environment. Acoustic and optical imaging technologies are conventional methods. However, the spatial resolution of acoustic images is pretty low, while light is severely scattered and absorbed by water medium, and optical images often degrade seriously [1] . One of the most effective solutions to above problems is to apply underwater laser imaging technology.The attenuation of blue-green light, whose spectrum wavelength is between 470 and 580 nm, is much less than that of light in other bands. Underwater laser imaging technology takes advantage of this phenomenon, adopts blue-green pulsed laser generator as an illuminator and range gated charge-coupled device (CCD) camera, effectively suppresses the backscattering and achieves a detection range of 4-6 times of conventional camera in intervening water medium [2] . However, due to absorption and scattering of water medium, as well as the laser pulse stretching, the characteristics in underwater laser images, such as speckle noise [3] , blur effect and non-uniform illumination [4] , have brought great difficulty for underwater laser image segmentation. Conventional segmentation algorithms, such as statistical, graph-based and morphological methods, show high sensitivity to these characteristics and always give inaccurate segmentation results. For instance, the fuzzy C-means and watershed methods are sensitive to speckle noise; the mean-shift method shows robustness against noise, but its performance is seriously decreased by non-uniform illumination. The uncertainty in weighted undirected graph (WUG) of norm...

Underwater image processing and analysis: A review

“…In this method, an iterative MapReduce procedure is designed to enhance the efficiency of the algorithm. In accordance with the principle of minimum entropy loss, Wang et al [50] devised a deep neural network framework in order to divide underwater images into several regions via a gradient optimization algorithm.…”

DPANet: Dual Pooling‐aggregated Attention Network for fish segmentation

“…In addition, Wang et al. [55] designed a deep neural network framework that divides underwater images into regions using a gradient optimisation algorithm.…”

“…Zhang et al [52], on the other hand, combined the fractal theory with the characteristics of underwater images and proposed a Brownian random field method that can effectively help in the segmentation of underwater images. In addition, Wang et al [55] designed a deep neural network framework that divides underwater images into regions using a gradient optimisation algorithm.…”