Super-Resolution Reconstruction And Detection Of Coral Reef Fish Images Based On Deep Learning

The surveillance image of underwater coral reef fish images contains a large number of coral reef fish species,as well as activities,such as location,species,quantity and other information,which can be used as a source of marine ecosystem and biodiversity detection.The super-resolution reconstruction and detection of underwater coral reef fish images could help marine biologists understand the natural underwater environment and study the behavior of different marine lifes and the interaction between them,and so it's crucial to analysis and study the coral reef fish data.With the realization of underwater video surveillance solutions,marine biologists can obtain a large number of surveillance images of coral reef waters at low cost.However,due to the difficulty of underwater environment shooting and the fact that the camera is underwater,the photos are affected by water flow and illumination.Although the underwater surveillance images continue to be produced in large numbers,the images of coral reef fishes produced are not able to reach the level that humans can appreciate,and have a modest impact on marine biologists' analysis and research about marine life and seabed ecology.There is an urgent need to improve the quality of images or videos of marine reef fish and to provide effective methods for automatic detection of coral reef fish,helping oceanographers to monitor and study marine life and marine ecology more effectively.At present,there are few researches on super-resolution methods for underwater surveillance images,and the performance of super-resolution results for coral reef fish images needs to be improved.The main work and achievements of this paper about the above two issues are as follows:First,we propose a high-resolution frame reconstruction method for single-frame coral reef fish images based on a step-by-step expansion strategy to further improve the effectiveness of super-resolution reconstruction of coral reef fish images.The method combines the shallow detail information of the coral reef fish image with the deep semantic information to enhance the feature reuse of the super-resolution reconstruction model,and adopt multi-step expansion to better learn the information required for each super-resolution expansion.The final reconstruction result has an average PSNR of 34 on a single frame image super resolution,5 points higher than the direct bilinear interpolation,3 points higher than the original super resolution reconstruction network,and the final reconstruction result has more details of the information.Secondly,we propose an end-to-end super-resolution reconstruction and detection framework for coral reef fish images based on target detection.Existing super-resolution reconstruction methods based on single-frame coral reef fish images do not effectively exploit the correlation of the image detail between adjacent frames.Therefore,we present an end-to-end deep learning framework combining generative adversarial network and target detection network for high-resolution images,focusing on the super-resolution reconstruction of coral reef fish images based on target detection and making the resulting image pay more attention to fish targets at the same time.The experimental results show that the multi-frame-based coral reef fish super-resolution method can effectively improve the quality of reconstructed images,and the image's perceptual index can be significantly reduced,thus it can generate super-resolution images more in line with human visual senses.The super-resolution reconstruction method of coral reef fish image based on target detection can also improve the final accuracy of image detection of the coral reef fish.For the original coral reef fish resolution detection results,the final detection MAP result can be improved by about 5%.