Laser & Optoelectronics Progress, Volume. 57, Issue 6, 060002(2020)
Review on Key Technologies of Target Exploration in Underwater Optical Images
Figures & Tables(5)
Table 1. Comparison of underwater image enhancement methods
View tableTable 1. Comparison of underwater image enhancement methods
Method Principle Characteristic Limitation Histogram-based Improving image contrast through nonlinear stretching of pixel values Bright color, real-time Lack of texture details, increasing noise Retinex-based Removing the influence of illuminating light from the image to obtain the reflective properties of the object Bright color, high contrast, more detail information Over-enhancement for the brighter patch, unsaturation for darker region Fusion-based Combining relevant information from two or more images into a single image, which is more informative than any of the inputs Effective color correction, high contrast Less robust, affected by artificial illumination Deep learning-based Imitating the working of the human brain in processing data, and improving the quality of underwater image through network training Effective color correction, robust contrast stretching Hard to get ground-truth images, time-consuming for training
Table 2. Comparison of underwater scene coefficient estimation methods
View tableTable 2. Comparison of underwater scene coefficient estimation methods
Method Improvement Characteristic DCP-based Tang et al.[ 25 ]:the depth-of-field map is obtained bydisparity between bright and dark channels, so as to estimate the background color of water body and transmittance map more accurately High real-time performance, more authentic colors, influence of artificial light sources, and inadequate robustness and adaptability Xu et al.[ 26 ]:estimation oftransmittance map of images by red channel algorithms Xie et al.[ 27 ]:using the relationship betweenthe wavelength of visible light and the scattering coefficient to obtain transmittance maps of each color channel Deep learning-based[ 23 ]URCNN: a convolutional layer and ReLU are used to generate feature maps, then the batch normalization is added between convolutional layer and ReLU to speed up training process. This pattern is repeated until the transmission map is output More authentic colors, robustness, and being time-consuming for training UIRNet:the transmission map and the background light are estimated and computed by BL-Net and TM-Net, respectively WaterGAN[ 24 ]:taking in-air images anddepth maps as input and generating corresponding synthetic underwater images as output
Table 3. Comparison of depth neural network for underwater target detection and recognition
View tableTable 3. Comparison of depth neural network for underwater target detection and recognition
Author Method Advantage and limitation Salman et al.[ 43 ]A hybrid approach involving GMM, optical flow, and deep R-CNN to fine-tune the categorization of fish Higher classification accuracy; requirement on relatively more computational resources Siddiqui et al.[ 44 ]A cross-layer pooling algorithm using a pre-trained convolutional neural network as a generalized feature detector No need for a large amount of training data; high classification accuracy; requirement of extensive computations Sun et al.[ 45 ]A CNN knowledge transfer framework for underwater object recognition and extracting discriminative features from relatively low contrast images High real-time performance; no need for a large amount of training data; lower robustness Chuang et al.[ 46 ]An underwater fish recognition framework that consists of a fully unsupervised feature learning technique and an error-resilient classifier Successfully handled data uncertainty and class imbalance in practical classification applications; lower robustness Cao et al.[ 42 ]A method to combine CNN and hand-designed features to improve classification performance Being robust, reducing feature dimensionality without decreasing classification performance; contour extraction needing original video frames
Table 4. Comparison of underwater target tracking methods
View tableTable 4. Comparison of underwater target tracking methods
Method Improvement Principle Characteristic Optical flow-based Modified Lucas-Kanade optical flow method based on pyramid hierarchy and affine transformation Estimating the position by calculating the velocity in two consecutive frames Need of a large number of precise image feature points Mean shift-based 1) Adaptive mean shift algorithm through color histogram based on background and target region; 2) adaptive mean shift algorithm combined with edge information; 3) combined with color, texture, HOG features, and deformable multi-core algorithm Using the color histogram of the target as the search feature, and iterating the mean shift vector continuously Being robust to range variations; being influenced by multi-objects CNT-based 1) Fast-CNT: improving the computing performance by selecting adaptive K value and omitting background filter; 2) improved Fast-CNT: extracting each region containing a moving target through Gaussian mixture model Tracking target through two-layer forward convolution networks learning image features No need for off-line training which is time-consuming and requires lots of training data
Tools
Get Citation
Copy Citation Text
Sen Lin, Ying Zhao. Review on Key Technologies of Target Exploration in Underwater Optical Images[J]. Laser & Optoelectronics Progress, 2020, 57(6): 060002
Paper Information
Category: Reviews
Received: Jul. 9, 2019
Accepted: Aug. 28, 2019
Published Online: Mar. 6, 2020
The Author Email: Zhao Ying (1460419171@qq.com)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20