Research On Deep Learning-based Underwater Target Detection Algorithm

Due to the speedy progress of the world economy and advancements in science and technology,people are increasingly exploring the ocean through various underwater activities,which accelerates the development and utilization of various underwater resources.Underwater robots serve as one of the main carriers for these activities.However,the complexity of the underwater environment often leads to poor detection performance for underwater robots.The objective of this paper is to propose a deep learning-based algorithm for underwater target detection,which is designed to tackle this problem.Aiming to address the challenges of blurred target features and difficulty in feature extraction caused by the complexity of the underwater environment,this document proposes a method based on the improved convolutional neural network YOLOv5 s for detecting targets in underwater environments.Firstly,expand the number of bottlenecks in the first C3 module to enhance the model’s feature extraction of blurred targets in the image;secondly,design two new C3 modules to replace the original C3 modules in the backbone network,so that the model’s detection capability is enhanced by incorporating positional information into the feature training process,while also decreasing the number of model parameters;ultimately,the SE attention unit is integrated into a particular location of the core network,so that the model can strengthen the attention to blurred objects in the image while suppressing the weight of unimportant objects.After various experimental verifications,compared with the original model,the model for detection suggested in this chapter advances the accuracy of detecting the particular submerged objective in the underwater setting while guaranteeing the swiftness of detection.After the backbone network extracts features at different levels,it is necessary to efficiently fuse these features using a neck network.To address this,this paper further proposes an improved underwater object detection algorithm based on the YOLOv5 s model with an enhanced neck network.First of all,the neck network of the YOLOv5 s model is improved by using the BiFPN(Bidirectional Feature Pyramid Network)structure,which improves the fusion ability of the neck network for different levels of feature maps.Subsequently,a section of the upsampling module in the neck network is enhanced by applying the CARAFE upsampling operator,and the receptive field of the upsampling module is improved,so that it can better utilize the spatial information of the feature map.Ultimately,the loss function utilized by the model and the activation function employed in the convolutional block are substituted.These upgrades have the potential to quicken the convergence speed of the model and augment the feature extraction proficiency of the convolutional block.By means of diverse experiments conducted on the model,it was demonstrated that the suggested algorithm is effective,as evidenced by the model’s parameter quantity,detection speed,and detection accuracy.