Sen Lin, Ying Zhao. Review on Key Technologies of Target Exploration in Underwater Optical Images[J]. Laser & Optoelectronics Progress, 2020, 57(6): 060002
Copy Citation Text
Hard to get ground-truthimages, time-consumingfor training
Table 1. Comparison of underwater image enhancement 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 toestimate the background color of water body andtransmittance map more accurately
High real-time performance,more authentic colors,influence of artificial lightsources, and inadequaterobustness and adaptability
Xu et al.[26]:estimation oftransmittance map of images byred channel algorithms
Xie et al.[27]:using the relationship betweenthe wavelength of visible light and the scatteringcoefficient to obtain transmittance maps of each color channel
Deeplearning-based[23]
URCNN: a convolutional layer and ReLU are used to generatefeature maps, then the batch normalization is addedbetween convolutional layer and ReLU to speed up training process.This pattern is repeated until the transmission map is output
More authentic colors,robustness, and beingtime-consuming for training
UIRNet:the transmission map and the backgroundlight are estimated and computedby BL-Net and TM-Net, respectively
WaterGAN[24]:taking in-air images anddepth maps as input and generatingcorresponding synthetic underwater images as output
Table 2. Comparison of underwater scene coefficient estimation methods
Author
Method
Advantage and limitation
Salmanet al.[43]
A hybrid approach involving GMM,optical flow, and deep R-CNN tofine-tune the categorization of fish
Higher classification accuracy;requirement on relatively morecomputational resources
Siddiquiet al.[44]
A cross-layer pooling algorithm usinga pre-trained convolutional neuralnetwork as a generalized feature detector
No need for a large amount of training data;high classification accuracy;requirement of extensive computations
Sunet al.[45]
A CNN knowledge transfer framework forunderwater object recognition andextracting discriminative featuresfrom relatively low contrast images
High real-time performance;no need for a large amount of training data;lower robustness
Chuanget al.[46]
An underwater fish recognition frameworkthat consists of a fully unsupervisedfeature learning technique and anerror-resilient classifier
Successfully handled data uncertaintyand class imbalance in practicalclassification applications;lower robustness
Caoet al.[42]
A method to combine CNN andhand-designed features toimprove classification performance
Being robust, reducing featuredimensionality without decreasingclassification performance;contour extraction needing original video frames
Table 3. Comparison of depth neural network for underwater target detection and recognition
Method
Improvement
Principle
Characteristic
Optical flow-based
Modified Lucas-Kanade opticalflow method based on pyramidhierarchy and affine transformation
Estimating the positionby calculating thevelocity in twoconsecutive frames
Need of a largenumber ofprecise imagefeature points
Mean shift-based
1) Adaptive mean shift algorithmthrough color histogram based onbackground and target region;2) adaptive mean shift algorithmcombined with edge information;3) combined with color, texture,HOG features, and deformablemulti-core algorithm
Using the colorhistogram of the targetas the searchfeature, and iteratingthe mean shiftvector continuously
Being robust torange variations;being influencedby multi-objects
CNT-based
1) Fast-CNT: improving thecomputing performance byselecting adaptive K value andomitting background filter;2) improved Fast-CNT:extracting each region containinga moving target through Gaussianmixture model
Tracking target throughtwo-layer forwardconvolution networkslearning image features
No need for off-linetraining which istime-consumingand requires lotsof training data
Table 4. Comparison of underwater target tracking methods