Research On Underwater Image Restoration And Target Recognition Detection Technologies

The ocean covers 71%of the surface of the earth,and contains rich natural resources,which is one of exploitable territories for the development of human.With the continuous exploration and development for the ocean,the detection system of underwater vehicle has gradually become research focus in the field of underwater detection.However,owning to the scattering and absorption effects on the light caused by water medium and fine particles,images acquired in underwater environment usually present poor visual quality.The resulting degradations in underwater images negatively affect the performance of several subsequent image analysis algorithms,such as image segmentation,aquatic organism recognition,organism detection and so on.The purpose of image restoration is to improve the quality of degraded underwater images,suppress the negative effect caused by low contrast,color distortion,noise and other defects to some extent,and to provide high-quality images for subsequent image analysis algorithms,which is of great significance to increase the discriminability of underwater targets.Accordingly,for an underwater exploration system,the underwater image restoration technology has become one of pivotal technologies.Moreover,in complex marine environment,the excellent perception ability is significant for underwater detectors to autonomously implement tasks.Hence,the technologies of underwater target recognition and detection are also other pivotal technologies for the underwater exploration system.Traditional underwater target recognition and detection methods rely on artificially designed features,but these features are incapable of accurately representing the high-level semantic information of targets,and thus perform poorly in terms of discriminability and generalization.Unlike the approach of artificially designing feature,the convolutional neural network（CNN）automatically learns image features from a large amount of data,and implements the feature transformation from low-level texture features to high-level abstract representations by flowing features from layer to layer.Owning to the powerful ability in terms of feature learning,CNN presents exceptional performance and generalization ability in targetrecognition and detection tasks.Nevertheless,due to the complex underwater environment,underwater images usually present poor visibility,weakening the discriminative representations.Moreover,underwater images cannot be readily captured,resulting in a scarcity in underwater image samples;hence,training an effective CNN for underwater organism detection or recognition is challenging.The diversity in underwater species further decreases the recognition accuracy or detection precision for underwater targets.Hence,improving the quality of underwater images and the performance of CNN in terms of feature transformation,so as to increase the recognition accuracy or detection precision,is research focus in the field of underwater exploration,which is significant for improving the autonomous work level of underwater detectors.Taking above-mentioned factors into consideration,the research theme of underwater image restoration and target recognition detection technologies is proposed.The research tasks and innovations are as follows:（1）To address the defects in underwater images,such as poor visibility and color distortion,a method for restoring underwater images is proposed by decomposing curves of attenuating color.Firstly,a novel scoring formula is designed by analyzing the characteristics of background regions in underwater images.With the designed scoring formula,a hierarchical searching approach using quad-tree is employed to accurately compute the background light.Subsequently,some prior information related to color attenuating curves of underwater image is explored,and initial transmission maps are achieved by decomposing these color attenuating curves.Moreover,to accurately restore underwater images,the mathematical relationship and geometric constraints on the RGB channels of transmission maps are applied to amend the initial transmission maps.Finally,to address unbalanced absorption effect caused by water medium on the light,a color compensation algorithm based on the bright channel is proposed to further balance the color of restored images.（2）To suppress noise in restored images,a novel approach based on robust underwater imaging model is introduced to restore underwater images.Firstly,the mechanism of amplifying noise in underwater image restoration is investigated,and a robust underwater imaging model is constructed based on the investigation.Secondly,to address the ill-posed problem caused by multiple unknowns in the robust underwater imaging model,a maximum posteriori probability method is proposed to decompose illumination,reflection and noise components,and two post-processing methods are employed to correct the brightness and reflection components to achieve restored underwater images.Finally,in existing restoration algorithms,all parameters that are set base on empirical methods are not guaranteed to be optimal;hence,a parameter optimization strategy based on genetic algorithm is exploited to optimize the parameters of the proposed restoration approach.（3）To address the issue that the static convolution kernels cannot dynamically adapt to varying input features,a novel inner feature and convolution kernel calibration module is designed to increase the recognition accuracy of underwater organism.Firstly,the performance that static convolution kernels aggregate features is analyzed,and a strategy of calibrating feature and convolution kernel is proposed based on the analysis.Secondly,the feature attention mechanism is studied to emphasize significant features in spatial and channel dimensions,and an approach of calibrating convolution kernels is explored to assign different convolution kernels to different features.Finally,by using the grouping convolution strategy,the convolution kernel and feature calibration operations are respectively performed in two branches to decrease the complexity of the module.The proposed module applies the strategy of split-transform-merge to implement different calibration operations in two parallel branches.Consequently,it is capable of not only achieving dynamic convolution,but also establishing dependencies of features in spatial and channel dimensions,so as to improve the ability of transforming features and recognition accuracy of underwater organism.（4）Different feature layers in one feature pyramid usually share one detection head,which weakens the feature transformation ability of the detection head.To address this issue,pyramid decoupling modules are designed to increase the detection precision of underwater target.Firstly,the issue caused by the connection mode that the feature pyramid shares one detection head is analyzed,and a feature pyramid decoupling strategy is proposed based on the analysis.Secondly,an approach of fusing convolution kernels is studied to learn a fusion factor from different layer indexes or features of the pyramid,used to perform feature pyramid decoupling.Finally,a target detection network is constructed by using an anchor-free detection network as basic framework to implement underwater target detection.Because the convolution kernel calibration factor is learned from different layer indexes or features of the pyramid,the proposed module can assign unique convolution kernels to different layers of the feature pyramid,so as to improve the feature transformation ability of the detection head and increase the target detection precision.