Research On Underwater Biological Object Detection Algorithm Based On Vision Transformer

With the rapid development of ocean observation technology,underwater target detection plays an increasingly important role in marine economic fields such as fisheries and aquaculture.It has broad application prospects in the development and utilization of marine resources and the promotion of marine economic development.Unlike the more mature general target detection technology,current underwater target detection technology introduces many problems due to the special nature of the background.Specifically,(1)due to the complex and diverse underwater application scenarios,the previous solutions require the use of multiple models to enhance the detection,which not only consumes a large amount of computing resources but also has poor results and affects downstream tasks.(2)The underwater organisms to be detected have the problems of multi-scale and clustered distribution in the image,which makes the detection difficult.(3)In actual engineering scenarios with limited computing resources,although single-stage object detection algorithms based on deep learning can meet real-time engineering requirements,their low detection accuracy brings difficulties to practical use,and their performance urgently needs to be further improved.To address these issues,the research content of this paper is as follows:(1)To address the challenge of a single model being unable to complete multiple image enhancement tasks,this paper proposes an image enhancement algorithm called ECO-GAN(Economic-Generative Adversarial Net)based on generative adversarial networks.Firstly,the extracted features are respectively given to three image enhancement branches with different functions,and a cross-stage feature fusion module is used to strengthen the connection between the branches.Secondly,to address the issue of the large calculation amount of the multi-functional model,the feature extraction network is reused to reduce the computation amount and improve the model’s inference speed.Based on experimental data,the results show that ECO-GAN can achieve blind enhancement of images without relying on prior information,and it improves the quality of the enhanced image objectively and subjectively.The peak signal-to-noise ratio(PSNR)and the structural similarity index(SSIM)of the enhanced image can reach 32.71 d B and 0.9873,respectively.(2)To address the difficulty of detecting underwater organisms due to their clustered distribution,as well as the adaptability of general detection models in underwater engineering backgrounds,this paper proposes a new object detection model based on the Swin Transformer under the assumption of sufficient computing power.Firstly,to tackle the problem of multi-scale and clustered distribution of organisms,the paper proposes the Selective Kernel Convolutional Network(SKANet).Then,to address the issue of single feature type extraction in the original Convolutional Neural Networks(CNN)structure,the paper proposes the Attention Fusion(ATF)module,which combines spatial attention,channel inter-attention,and self-attention fusion.Next,to address the problem of information asymmetry between training and testing phases in object detection models,the paper proposes a detection head that integrates classification and location information and recalculates the confidence score of the target based on its position and classification information.Finally,the proposed model is validated using the URPC dataset.Experimental results based on measured data show that the proposed model improves the average detection accuracy by 3.2% compared to the baseline model.(3)In order to improve the detection accuracy of a single-stage model for real-time underwater object detection in scenarios with limited computing resources,this paper proposes a new YOLOv5-based single-stage underwater object detection model.Compared to the original YOLOv5 model,this model includes five improvements adapted to the underwater engineering background: 1)improved feature extraction module in the detection model with model fine-tuning;2)two training strategies proposed from the perspective of data augmentation;3)a feature extraction module that fuses YOLO and Vi T to address the problem of single feature extraction type around the original CNN structure;4)a small object random enlargement strategy to address the difficulty of small object detection in the dataset;5)a rare sample no overlap filling strategy to address the problem of sample imbalance in the underwater dataset.Experiments based on the URPC dataset show that the improved YOLOv5 model has an average detection accuracy improvement of 3.06% compared to the original model.It is worth noting that after adopting the data augmentation strategy,the model can achieve a detection accuracy of 92.49% when the m AP threshold is set to 0.5.This result demonstrates the superiority of the proposed model and the effectiveness of the data augmentation strategy.