Research On Underwater Imaging Model-based Image Restoration And Algorithm Acceleration

Underwater images play a vital role in marine scientific research and intelligent perception of underwater information.However,underwater images often have degradation phenomena such as color distortion and low illumination,which hinder underwater vision tasks.In addition,some underwater image restoration algorithms have high computational complexity and space complexity.In particular,the deployment of deep learning-based methods in underwater robots is limited by power consumption and processor computing performance,which affects practical applications.Therefore,research on underwater image restoration and its algorithm acceleration has high scientific significance and practical value.The main contents of this thesis are as follows:(1)An underwater image generation method based on an imaging model is proposed,and more than 20,000 pairs of underwater images that simulate different water types and imaging conditions are simulated to solve the lack of paired data sets in the current underwater image restoration research field.Comparative experiments show that the proposed generation method can effectively simulate underwater scenes in terms of scattering and color distortion degradation.(2)An underwater image restoration network based on underwater imaging model is proposed.First,the network consists of an encoder-decoder structure.The encoder network extracts the feature information of the image,and the decoder networks respectively estimate the underwater imaging parameters and then reconstruct the restored image through the imaging model.Secondly,the corresponding network module and loss function are designed for the restoration task.Finally,the evaluation of multiple dimensions verifies that the proposed underwater image restoration network has better restoration performance and higher generalization,and is superior to comparison algorithms in some objective evaluation indicators.(3)Realize the acceleration and application of underwater image restoration algorithm.First,the original network is simplified and the module is lightened,and then the model is accelerated by combining model pruning and quantization strategies.Experiments show that the acceleration method effectively compresses the number of parameters and calculations,and the accelerated network has a higher model inference speed while ensuring the accuracy of use.In addition,the deployment of mobile terminal equipment is also realized,and the inference speed meets the needs of real-time processing.Finally,the restoration network is applied to a variety of underwater vision tasks,and its value to underwater tasks is verified.