Research On Underwater Target Detection Algorithm Based On Improved YOLOv5S

Due to the needs of human development and the progress of marine information technology,the rapid development of marine pastures has been promoted,and the scale of artificially cultured seafood has been expanding.Underwater target detection technology is the core link to realize automatic monitoring and fishing technology of seafood.It is of great significance to improve the efficiency of seafood detection,reduce the risk of manual operation and improve the fishing environment of seafood.However,the underwater imaging environment is complex and the underwater image quality is relatively low,which makes the underwater target detection more difficult.In this paper,four common marine biological targets,holothurian,starfish,echinus and scallop,are taken as research objects.Combining underwater image preprocessing and computer vision technology,using multi-scale weighted fusion CLAHE underwater image enhancement method,an improved underwater target detection algorithm based on YOLOv5S is designed,and the hardware deployment experiment of the improved underwater target detection algorithm mobile platform based on YOLOv5S is completed.First of all,in this paper,CLAHE image enhancement algorithm is used to improve the quality of the dataset image,to solve the underwater image distortion,blur and other issues that affect the underwater target detection effect.Through the analysis of underwater imaging environment,the mechanism of ACE,CLAHE and RCP algorithms to improve the quality of ocean images is expounded.The characteristics of various preprocessing algorithms are summarized,and a multi-scale weighted fusion CLAHE underwater image enhancement algorithm is designed for the phenomenon of image detail loss caused by CLAHE image enhancement algorithm.Based on different image processing algorithms,the comparison of visual effects,information entropy,mean value,variance,average gradient and other indicators after image processing,combined with the training effect of the model,shows the superiority of the multi-scale weighted fusion CLAHE algorithm adopted in this paper.Secondly,on the basis of YOLOv5S target detection algorithm,an improved YOLOv5S underwater target detection algorithm is proposed,which realizes the effective detection of marine biological targets.Aiming at the problem of missing detection when detecting overlapping occluded objects,Soft-NMS is used instead of NMS;Introducing Coordinate Attention Mechanism into YOLOv5S Model,The ASFF module is added before the output layer,which makes the features of different scales get better fusion effect,so that the model can get better features and optimize the detection effect of the model.Simulation results show that the improved YOLOv5S underwater target detection model can greatly improve the detection accuracy without losing speed.In addition,several target detection models such as YOLOv3,YOLOv4,YOLOv5S,YOLOv5M and improved YOLOv5S are compared and simulated,which shows the effectiveness of the improved method.Finally,the improved YOLOv5S underwater target detection model is deployed on the development platform of Jetson Xavier NX,and the comparative evaluation experiment is carried out.In the implementation of deployment,Deepstream deployment tool is used to optimize the model.In order to achieve faster model reasoning speed,TensorRT is used to accelerate the improved YOLOv5S model,and INT8 is used to quantify the model,so that the detection speed of the model is further improved.Compared with the experimental results,the speed of the accelerated model reasoning has been greatly improved,better meet the real-time requirements.