Research On Object Detection Of Lightweight Seafood Based On YOLOv5

Seafood refers to precious and highly nutritious food ingredients that are unique to the ocean,such as seacucumbers,seaurchins,scallops,abalone,and oysters,among others.These ingredients are highly valued in both domestic and international markets due to their unique taste and rich nutritional value.With the increasing demand for seafood object detection model suitable for underwater robot and other embedded equipment for the intelligent construction of Marine pasture.In recent years,with the continuous development of computer vision and deep learning technology,using machine learning methods for automated detection and classification of seafood has become a research hotspot.These methods can effectively improve the accuracy and efficiency of detection,while reducing labor costs,and have broad application prospects.However,the current deep learning methods are highly accurate but also complex,making it difficult to run on embedded devices.In this study,we will use YOLOv5 as the benchmark network and make lightweight improvements to it.We will also use knowledge distillation to improve the model’s detection accuracy.The improved algorithm not only achieves significant lightweighting but also maintains the original detection accuracy.The main work of this paper is as follows:(1)In this paper,we propose a lightweight model,FRC-YOLOv5.We make lightweight improvements to the three components of YOLOv5: backbone,neck,and head.Firstly,for the backbone,we use four lightweight design principles and a structure similar to Shuffle Net V2,and further reduce its weight by using the CSPNet method and Rep Ghost to improve the inference speed.Secondly,for the neck,we use DW convolution and a unified output channel number to reduce the number of parameters,and enrich the intermediate layer information using weighted cross-layer connections.Finally,for the head,we decouple the output and use EIOU as the bounding box loss function.Experimental results show that compared to YOLOv5 s,FRCYOLOv5 has a significantly reduced computational complexity,only 12.1% of that of YOLOv5 s,and a reduced parameter count by 5M.The inference speed on GPU has increased by 44 fps,and on CPU it has reached 31 fps,more than twice that of YOLOv5 s.The detection accuracy,as measured by m AP@.5,is 79.2%,which is 9.6% lower than YOLOv5 but higher than other lightweight models.(2)A knowledge distillation scheme for sea delicacy detection based on a hybrid of local attention mechanism and global attention mechanism has been proposed.YOLOv5 s was used as the teacher network,FRC-YOLOv5 as the student network,and spatial attention mechanism and channel attention mechanism were used as local knowledge distillation,while non-local attention.mechanism was used as global information distillation.The experimental results show that the FRC-YOLOv5 after knowledge distillation has the same detection speed as before,but the m AP of the distilled FRC-YOLOv5 reached 86.1%,which is 6.9 percentage points higher than before,and better than other knowledge distillation schemes.