Research On Underwater Image Superresolution Reconstruction Based On Deep Learning

In recent years,with the growth of population and the acceleration of consumption of land resources,the ocean has become the focus of future resource development.The ocean covers most of the earth’s surface area and is rich in biological and petroleum resources,attracting more and more researchers to participate in the development and utilization of marine resources.However,considering the tolerance of the human body under the sea,underwater robots or underwater vehicles have become the most suitable visual information communication medium between researchers and the ocean.Different from the imaging process of terrestrial natural images,due to the complexity of the underwater environment and the limitations of physical equipment,the underwater images acquired by underwater vehicles often face problems such as blurred images and low resolution.Good perception of underwater scenes.In this paper,a method based on deep learning is used to study the loss of high-frequency details and high model redundancy in the process of super-resolution reconstruction.The main research contributions are summarized as follows:(1)In order to further effectively distinguish the features to be refined and the features that need to be transferred across layers.This paper proposes a multi-stage information distillation and spatial attention mechanism for underwater image superresolution reconstruction method SRIDM.Based on the ordinary residual network,the algorithm introduces a global feature fusion structure,an information distillation mechanism and a spatial attention module,which further improves the feature expression ability of the model.A global feature fusion structure is used to enhance information flow and feature reuse between layers.The designed information distillation structure extracts image features step by step,and puts more learning tasks of the model on the reconstruction of image texture details;the spatial attention module enables the network to adaptively assign weights to the image space,and more attention The focus is on areas carrying high-frequency information to help the network recover high-frequency details quickly.Through model ablation experiments,both the global feature fusion structure and the spatial attention module adopted in this paper can effectively improve the quality of the reconstructed image,and the improvement effect is better when the two are coupled.The best performance is obtained with the model distillation rate set to 0.25.The comparative experiments on the USR-248 test set show that the images restored by this method are superior to other comparison algorithms in terms of subjective visual effect and objective evaluation quality.When the amplification factor is 4,the peak signal-tonoise ratio and structural similarity reach 27.7640 d B and 0.7640,respectively.In addition,the algorithm in this paper is also a lightweight model,which greatly reduces the amount of model parameters and computational complexity while maintaining performance.Through the underwater small target detection experiment,the results show that the superresolution method proposed in this paper also has a certain improvement effect on the underwater small target detection task,and has a significant improvement effect in terms of prediction accuracy and positioning accuracy.In order to further reduce the forward reasoning time,this paper then adopts the structure reparameterization technique to optimize and improve the RFDB module,and obtain the accelerated reasoning version SRIDM+ model.Compared with SRIDM,this model not only improves the inference speed of a single image,but more importantly,thanks to the performance advantages brought by the multi-way branch structure in the training phase,the PSNR value of SRIDM+ on the same test set Get significant gains.(2)In order to provide some convenience for researchers in related fields,this paper develops a simple and practical underwater image super-resolution reconstruction system,which integrates image super-resolution and small target detection functions,and provides certain auxiliary functions.For example: image detail viewing,histogram equalization,reconstructed image quality assessment,etc.,to help users get a better experience.At the architectural level,the system is divided into four modules: GUI user interface display module,business logic control module,back-end algorithm module and data persistent storage module.Functional decoupling and independent packaging of algorithms facilitate future system maintenance and function expansion.Super-resolution algorithm selection parameters provide 7 deep learning-based methods including SRIDM and SRIDM+,and configure four amplification factors of x2,x3,x4,and x8 for computer vision downstream tasks(image classification,object detection,semantic segmentation,etc.)to provide reconstructed images that meet different super-resolution effects.The effectiveness and practicability of the underwater image processing system developed in this paper in the engineering application of super-resolution algorithm and small target detection are verified by the display effect of each module.