The Algorithm Design And System Implement Of Sea Cucumber Detection Based On Deep Learning

Sea cucumber is not only one of the eight treasures,but also a valuable medicinal material.However,sea cucumbers mainly rely on divers for underwater fishing,with high work intensity,low efficiency,and a major threat to personal safety.Therefore,the research and development of sea cucumber automatic fishing equipment is an inevitable trend for the development of smart fisheries.The ability to accurately detect underwater sea cucumber targets in real time is the key to the development of automatic fishing equipment.The complex and variable underwater environment brings great challenge to the sea cucumber detection algorithms.This article takes underwater sea cucumbers as the research object,based on the deep learning technology to study the sea cucumber detection algorithm with high precision and low time consumption.The main work completed in this paper is as follows:(1)Aiming at the problem that the Faster R-CNN sea cucumber detection algorithm based on ResNet50 has low inference speed and cannot meet the practical application,a Balanced R-CNN fast sea cucumber detection algorithm is proposed,which guarantees high detection accuracy and achieves high detection speed.This algorithm first improves the Faster R-CNN detection structure and eliminates a large number of repeated convolution calculations.Second,though convolution features fusion and spatial attention mechanisms to generate features with rich details and strong robustness.This feature map is beneficial to improve the detection accuracy.Finally,the balance loss function is introduced during the network training process,which increases the gradient contribution of easy samples and speeds up model convergence.On the sea cucumber data set,the designed algorithm has a detection accuracy(mAP)of 87.49%.On the Faster R-CNN,the mAP value increases by 1.1,and the detection speed reaches 27 FPS.(2)Aiming at the problems of low detection accuracy and high false detection rate of the sea cucumber detection algorithm based on YOLOv2,a YOLOv2-TDM sea cucumber detection algorithm is proposed without sacrificing the inference speed.The algorithm first designes the TDM feature enhancement mechanism to expand the semantic information of the underlying features.Then we design the downsampling mechanism to reduce the size of the feature map,reduce the amount of calculation and speed up the inference speed.Finally we introduce a bottleneck block,which enlarge the receptive field of small objects and it doesn't reducing the resolution of the feature map.The experimental results show that the proposed algorithm has a mAP value of 88.21% on the sea cucumber data set,an average detection precision(AP)is increased by 6.52% compared to YOLOv2,and it takes only 11 ms to detect an image.(3)An underwater sea cucumber target detection system platform is established,which integrate the software and hardware platforms of the underwater robot and the object detection software program.Experiments on sea cucumber target detection in image and video files verify the feasibility and effectiveness of the proposed algorithm for underwater sea cucumber target detection.In addition,the field test experiments of the pool verify whether the detection system can quickly and accurately detect sea cucumber targets in the underwater environment.