Research On Super-resolution Algorithm For Underwater Images Based On Improved Channel Attention Residual Network

The Underwater image super-resolution reconstruction technology is becoming increasingly important as the demand for high-quality underwater images grows in various fields,including underwater robotics,marine science research,marine resource exploration,aquaculture,marine environmental protection,and underwater archaeology.In these fields,underwater imaging equipment is widely used to capture images of the underwater environment,which can then be used for a range of purposes such as mapping the seabed,monitoring the health of marine life,and identifying potential hazards.However,the quality of underwater images tends to be relatively poor due to factors such as water scattering,absorption,and turbidity,which can lead to reduced resolution and prominent color distortion.Furthermore,the design,cost,and unique underwater environment of imaging equipment make it difficult to achieve ideal high resolution and clarity.To address this challenge,researchers have been developing various methods for underwater image superresolution reconstruction.One approach that has shown promise is deep learning-based methods,which have been successful in a wide range of image processing tasks,including super-resolution.Deep learning-based methods are particularly well-suited to this task because they can learn complex mapping functions between low-resolution and highresolution images,which are difficult to define using traditional methods.In this paper,we propose two methods for improving underwater image super-resolution methods and related image training and testing sets.The first proposed method is an improved channel attention residual network image super-resolution method.The second proposed method is a scaleaware underwater image super-resolution method based on meta-transfer learning.The improved channel attention residual network image super-resolution method builds on the existing residual network-based image super-resolution methods.Residual networks have been shown to be effective in image super-resolution because they can learn the residual mapping between low-resolution and high-resolution images.Our method introduces an improved channel attention residual module,which is designed to enhance the performance of the network.Specifically,we introduce a multi-spectral channel attention mechanism that improves the network’s ability to selectively attend to different spectral channels of the input image.This mechanism is based on the observation that different spectral channels of an image may have different information content and therefore require different levels of attention.We optimize the structure of the channel attention residual module and validate the proposed network on multiple image super-resolution test sets,including Set5,Set14,B100,Urban100,and Manga109.The experimental results demonstrate the excellent performance of the proposed network,proving the effectiveness of the method.The second proposed method is a scale-aware underwater image super-resolution method based on meta-transfer learning.This method builds on the improved channel attention residual image super-resolution network and adopts the meta-transfer learning approach,which allows the general image super-resolution method to quickly adapt to underwater image tasks.Meta-transfer learning is a technique that transfers knowledge learned from one task to another,allowing the network to learn more quickly and efficiently.Our method introduces a scale-aware module,which is designed to improve the up-sampling method for image reconstruction and enable the network to complete image super-resolution tasks of arbitrary scale.We train the model on the publicly available underwater image superresolution training set UFO-120 and validate it on the public underwater image superresolution test set UFO-120.The experimental results show that the introduction of the metatransfer learning module improves the adaptability of the model,and the proposed network achieves excellent subjective and objective performance,proving the effectiveness of the proposed method.