Research On The Hardware Architecture Of Underwater Image Enhancement Algorithm Based On MBD

The work in this thesis originated from the first deep sea live streaming mission of China's first 10,000 m deep dive in the Mariana Trench.Due to the weak illumination,strong scattering effects and inconsistent light attenuation rates by different wavelengths in the underwater environment,the underwater images suffer from low contrast,blurred edges and color deviation.This seriously affects the quality of underwater video,as well as the effectiveness and accuracy of subsequent high-level image information extraction.Both domestic and foreign researches on algorithms for underwater image quality enhancement have made some progress,but there still has the problem of weak scene adaptation.In recent years,underwater image quality enhancement algorithms based on deep learning have made substantial improvement in performance.However,the deep learning algorithms require too much computational resources and cannot adapt to the demands of real-time processing in practical applications.To address the above problems,this thesis has a research on the improvement of underwater image quality enhancement algorithms by considering the requirements of live streaming mission and designs two lightweight deep neural networks based on multi-modal fusion.This thesis uses the MBD methodology to design modular hardware architecture by Matlab Simulink and finally completes the hardware deployment of deep network algorithms under resource-constrained conditions.The details and innovation points of this thesis are as follows:(1)To address the problem of poor scene adaptation of traditional underwater image quality enhancement algorithms,this thesis uses a combination of depth network and image fusion to design a multi-modal fusion-based underwater image enhancement algorithm based on depth network.The depth network structure in this thesis enhances the ability of the algorithm to adapt to different scenes and improves the generalization performance of the algorithm;multi-modal fusion can effectively take advantage of each modal and avoid influencing each other,which is conducive to obtaining the best image enhancement results.Inspired by the Water Net algorithm,this thesis augments the image data of the lightweight networks by adding a white balance module,a contrast enhancement module,a denoising module and an edge enhancement module,which significantly reduces the computational complexity of the algorithm while keeping the quality of enhanced images.(2)To address the problem of the inconvenient hardware deployment caused by the computational overhead,this thesis re-designs the existing network structures and proposes two lightweight network structures,Shuffle Water Net and Mobile Water Net,based on the idea of local feature extraction and global feature reorganization,significantly reduce computational complexity.It can be proved by the experiment that both network structures can improve the image quality with different hardware resource utilization.(3)For the needs of real-time processing applications,this thesis researches the hardware structures for the two lightweight networks.Through the resource scheduling strategy and parallel flow scheduling of algorithms,the resource utilization and calculating speed of algorithms are balanced,so that the algorithms can run on hardware platforms with the best performance.Considering the similar heterogeneity of deep network algorithms,the hardware architecture design method of deep network models developed in this thesis can be extended to different deep network models and provide a theoretical basis for deep network hardware deployment.(4)In order to solve the problem of inefficient iteration of algorithm model improvement,this thesis introduces the MBD design idea and uses Matlab Simulink tool to design the hardware architecture of the proposed algorithms,and successfully completes the fast iteration from algorithm design to hardware architecture design.The actual algorithm porting process proves that the MBD-based design approach can significantly accelerate the hardware architecture design process and achieve high execution efficiency during the improvement iteration of different deep network structures.