Research On Underwater Target Detection Algorithm Based On Improved YOLOv5

As the development of social economy continues,coastal countries are increasingly exploiting their waters,and conflicts between nations over water and island resources are intensifying.China,as a coastal country,has experienced continuous disputes over island ownership.Thus,the use of object detection for real-time and accurate acquisition of water surface information is crucial for ensuring maritime safety.In recent years,although significant progress has been made in the field of object detection,challenges such as water surface fluctuations,changes in lighting,occlusion,overlapping,and small target sizes still affect the performance of maritime object detection tasks.These factors can lead to low image quality and loss of feature information during the convolution and downsampling processes,consequently affecting detection accuracy and speed.To address these issues,this thesis applies deep learning techniques to improve the YOLOv5 model for maritime object detection tasks.The main contributions are as follows:(1)To address the difficulty of detecting small objects and low confidence scores,we introduce recursive gated convolutions and P6 detection heads,proposing a target detection algorithm based on recursive gated convolutions.First,add a P6 detection head to increase the size of the input image,in order to learn more detailed information and improve the accuracy of small object detection.Next,we replace the standard convolutions in the bottom-up path of the feature aggregation network with recursive gated convolution modules,efficiently implementing high-order spatial interactions that standard convolutions cannot achieve and improving model performance and confidence scores.Finally,compared to the original YOLOv5 s model,the improved model increases detection accuracy by 3.9%and recall rate by 3.1%,reaching 68.5% and 67.3%,respectively.(2)To address the poor detection performance of overlapping objects in the YOLOv5 model,we introduce the CA attention mechanism and the SIo U loss function,proposing a target detection algorithm based on the CA attention mechanism.First,we incorporate the CA attention mechanism into the backbone network,embedding positional information into the channel attention and enabling the network to extract a broader range of feature information while reducing parameter count and computational overhead.Next,we replace the CIo U loss function with the SIo U loss function to better calculate regression loss,more accurately locate targets,and improve network convergence speed.This improves the detection performance of overlapping objects.Finally,compared to the original YOLOv5 s model,the improved model increases detection accuracy by 2.2% and recall rate by 2.1%,reaching 66.8% and 66.3%,respectively.(3)To simultaneously address the difficulties of detecting small objects,low confidence scores,and poor detection performance of overlapping objects,we propose a YOLOv5 s fusion algorithm that combines the two aforementioned improvements.On the ABOships dataset,the YOLOv5 s fusion algorithm model increases detection accuracy by4.6% and recall rate by 3.4% compared to the original YOLOv5 s model,reaching 69.2%and 67.6%,respectively.