Research On Underwater Object Detection Algorithm

The underwater target detection algorithm is an extension and extension of general target detection in underwater scenes.It is an important research topic and plays a crucial role in the fields of underwater mining and seabed exploration automation.At present,general target detection algorithms based on deep learning have good solutions in conventional fields such as face detection and vehicle detection,and have achieved excellent results in terms of accuracy and speed.However,in the field of underwater special target detection,due to many small underwater targets,poor lighting conditions,mutual occlusion of objects,or blurred pictures caused by the movement of underwater shooting equipment,the underwater target detection scene is compared with general-purpose targets.Target detection is more difficult to capture the location of the target,and at the same time,due to the harsh underwater environment,image acquisition is more difficult,resulting in the lack of a large number of labeled datasets for large-scale supervised training.The number of pictures is small and the hardware resources are limited,so it is still necessary to continue study and explore how to construct an underwater target detection algorithm with higher accuracy and lower false detection.In this paper,a language-assisted model is proposed to assist the training of visual target detection,and it is used to provide text dimension information for training pictures with the text prompt data augmentation proposed in the paper.The paper designs the language-assisted model as a removable branch,extracts text information during training to prompt and correct the location and classification of objects for visual target detection under complex scene pictures,and removes the language-assisted branch during the testing process,thereby increasing only 15%training time,does not increase the computational loss of model inference,and improves the accuracy.At the same time,this paper also uses more data enhancement methods and GAN data generation methods to expand underwater pictures,and design a language-assisted loss function with a language-assisted model to balance the training direction of target detection.Finally,through the design of ablation experiments,this paper conducts a large number of experiments on the complex and large general target detection datasets such as Objects365,MS-COCO and OpenImages and the URPC series of underwater special target detection datasets and achieves significant improvement.The validity and generality of the algorithm in this paper are verified from the perspective.