Research On Marine Target Recognition Method Based On Deep Learning

As the treasure house of nature,the ocean is rich in resources.In recent years,the offshore fishery is facing a huge crisis due to over-exploitation.In order to curb the decline of fishery resources and enable sustainable development of fisheries,the country strongly advocates the construction of marine pastures.Due to the unpredictability of the marine environment,manual inspection of the marine pasture farming situation is characterized by high risk and low efficiency.With the development of machine vision technology,it has become a trend to use underwater robots to carry out inspection operations in marine pastures.Compared with the terrestrial environment,the marine environment is more complex,and the collected underwater images has the problems such as color shift,low resolution and many disturbances.The current deep learning target detection algorithms with large number of model parameters,slow detection speed and low efficiency are not fully suitable for the detection of biological targets in marine pastures.To address the above problems,this paper use image enhancement and improved lightweight YOLOv5(You Only Look Once version 5)deep learning algorithm for target recognition of marine organisms.The main research contents of are as follows:(1)To address the problems of blue-green bias,low illumination and blurred images of the collected marine biological image,this paper uses UGAN(Underwater Generative Adversarial Network)image enhancement algorithm to pre-process underwater images,and provide clear datasets for subsequent target detection model.Then the dataset is annotated,the experimental environment is configured and the training parameters are set,and the performance evaluation index of model detection is introduced.(2)The quality dataset processed by the enhancement algorithm is input to YOLOv5 s deep learning target detection algorithm for training.By comparing and analyzing the training results of different models,the YOLOv5 s algorithm has higher detection speed and better detection accuracy.The effectiveness of the UGAN enhancement algorithm on marine biological target recognition results is verified,and the detection accuracy is improved.It shows that the YOLOv5 s algorithm based on image enhancement for marine biological target detection has better results.(3)In order to improve the detection speed of the model and make the YOLOv5 s model embed into the low arithmetic hardware device,the lightweight YOLOv5 s algorithm based on the attention mechanism is used to recognize the marine biological targets.The lightweight model of Mobile Netv3 reduces the size of the model,decreases the number of parameters and computation,which sacrifices less detection accuracy thus leading to an increase in the detection speed.The attention mechanism is introduced to strengthen feature information to further improve the detection accuracy of the model,and the NAM(Normalization Attention Module)improves the detection of occluded targets.Experiments show that the M-NYOLOv5 s algorithm has improved the detection speed and detection accuracy of marine biological targets.(4)The M-NYOLOv5 s EH algorithm is proposed for marine biological targets with different sizes and small targets with missed detection.Using EIOU(EfficientIntersection Over Union)loss instead of CIOU(Complete-Intersection Over Union)of the original algorithm,the matching mechanism of the prediction frame and detection frame is optimized,and the detection accuracy is improved to some extent.Then the Head module is optimized to increase the small target prediction head,and the improved algorithm has an improved effect on the missed marine small targets.The improved MNYOLOv5 s EH algorithm can balance the speed and accuracy of marine biological target detection,while meeting the needs of hardware deployment and recognition.