Research On Scene Feature Fusion Detection And Lightweight Detection For Holothurian

With the increasing demand for holothurians,exploring holothurian target detection technology can assist robots to complete holothurian fishing tasks,which can help alleviate the dangers and risks of manual holothurian fishing operations,and can also promote sustainable development and utilization of marine resources.In recent years,research on land target detection algorithms has achieved remarkable results.For example,some remote sensing image analysis methods based on target detection have been able to automatically detect and classify different types of land,such as cities,farmland,and forests,and provide new insights into ecological protection,land management and Applications in areas such as urban planning provide strong support.Although the rapid development of computer vision and deep learning technology has brought great progress to land target detection,the research on underwater target detection algorithms is relatively weak,especially there are very few targeted holothurian detection algorithms.The research on holothurian detection algorithm mainly has the following difficulties:due to factors such as the complex seabed environment and the limitations of imaging equipment,underwater images have problems such as noise pollution,low contrast,and color distortion,which makes holothurian detection face the challenge of underwater target detection.Common difficulties;in addition,holothurians have an excellent selfprotection mechanism,and their body color can change with the color of the environment,so the physical characteristics of holothurians are highly similar to the environmental characteristics;this is also the biggest difficulty in holothurian detection tasks;in practice When the underwater target detection algorithm is applied,it is often faced with a shortage of computing resources for micro-embedded devices,which cannot be deployed and detected well.In view of the above problems,the main research contents of this paper are as follows:1.The research and development of underwater holothurian target detection and lightweight detection models at home and abroad are introduced in detail,laying a solid foundation for the follow-up experimental work.In addition,the basic composition and working principle of convolutional neural networks were briefly summarized,as well as the theoretical methods of two detection routes based on Anchor and Anchor free.Finally,the evaluation indicators for target detection were briefly introduced,laying the foundation for further research on underwater holothurian target detection calculation methods and evaluating algorithm performance.2.In view of common problems such as noise pollution,low contrast,and color distortion in underwater images,as well as the characteristics of blurred shape,high similarity to the background,and coexistence of special ecological scenes in holothurian recognition,this paper proposes a method based on improved Center Net and scene Feature fusion underwater holothurian target detection algorithm FA-Center Net.The algorithm uses Efficient Net-B3 as the backbone network to reduce the Params and FLOPs of the model,and increase the depth and width of the model to improve accuracy.Then,this paper designs the FPT combination module to fully focus and mine the ecological scene information of holothurians in different spaces and scales(for example,holothurian thorns,reefs,and aquatic plants often exist in the same scene with holothurians).The algorithm has high detection accuracy,can deal with fuzzy holothurians,small-sized holothurians and holothurian detection tasks in dense scenes,and achieves a balance between detection accuracy,Params and FLOPs.3.Aiming at the unbalanced comprehensive performance of the current deep learning single-stage detection algorithm and the difficulty of deployment on embedded devices,this chapter proposes a high-performance holothurian target detection algorithm ESGC-YOLOv5 s for embedded platforms.Based on the research data in the previous chapter,the algorithm uses YOLOv5 s with high precision and outstanding lightweight performance as the basic network,and has made more lightweight improvements and optimizations.First,ESNet,the latest research result developed by Baidu,is used to replace CSPDarknet as a lightweight and high-precision backbone network,and the backbone is lightweight.Quantization processing.These two strategies allow the algorithm to significantly reduce the Params and FLOPs of the model while only losing a small part of the accuracy,and obtain better lightweight performance.Experimental results show that compared with other underwater object detection methods,the algorithm proposed in this chapter has significant advantages in Params and FLOPs,while also ensuring considerable accuracy.The ESGC-YOLOv5 s algorithm always revolves around lightweight technology,which is suitable for mobile terminal deployment with extremely low computing power,and has practical application significance.