| The effective management of fishing vessels is the basis for ensuring the smooth progress of fishery activities.Through in-depth learning technology,the supervision ability of fishing vessels in the summer fishing moratorium can be effectively strengthened,which is of great research significance to the management and control of fishing vessels,the governance of fishing ports and the sustainable development of fishery resources.In the process of traditional methods for identifying ships in video,affected by factors such as technical architecture,dynamic background,hydrology,waves and camera jitter,there are problems of low recognition accuracy and high false alarm rate and missing alarm rate,resulting in video recognition can not well assist fishing port management and service.The development of deep learning technology provides technical support for solving this problem.The automatic feature extraction and strong generalization of deep learning technology can improve the recognition accuracy and robustness of fishing vessels in monitoring.Through the research of two-stage target detection algorithm Fast-R-CNN and single-stage target detection algorithm YOLOv5,this paper completes the production of fishing vessel data set,fishing vessel detection and the improvement design of existing algorithms.The specific work is as follows:（1）A fishing vessel target detection model based on improved Fast-R-CNN algorithm is proposed.Firstly,the anchor frame is optimized,and the size and position of the anchor frame generated by regional suggestion network（RPN）are optimized by using Kmeans++ clustering algorithm;Secondly,the vgg16 network in the original algorithm is replaced by the resnet50 network with better feature extraction,and the coordinate attention（CA）mechanism is added to the convolution layer of resnet50 network to strengthen the feature extraction;Then,FPN network is used to fuse the low-level feature map with more feature details with the high-level feature map with rich semantic information;Finally,in the prediction frame screening,the soft NMS method is used to replace the traditional NMS method to reduce the missed detection rate of overlapping targets.（2）A series of improved experiments are designed based on YOLOv5.Firstly,the anchor frame is re clustered by Kmeans++ algorithm to select the anchor frame size suitable for the fishing vessel data set;Then,CBAM attention mechanism is integrated into the backbone network of YOLOv5 to obtain more detailed features;Then,the weighted bidirectional feature pyramid network（BiFPN）is used to replace the original FPN + pixel aggregation network（PAN）structure for fast multi-scale feature fusion;Finally,the detection scale of large targets is removed,the detection scale of smaller targets is added,and three new detection scales are used to improve the detection accuracy of the model for small target fishing vessels.（3）The advantages and disadvantages of the improved Fast-R-CNN algorithm and the improved YOLOv5 algorithm are comprehensively discussed and analyzed from the perspectives of accuracy,recall,detection speed and detection robustness.In order to detect the robustness of the algorithm,the performance of the two algorithms under different sea conditions such as heavy fog environment,incomplete display of fishing boat hull,different orientation of hull and influence of illumination are compared and analyzed,and the improvement ideas are put forward according to the shortcomings of the algorithm.（4）A fishing vessel identification system based on improved YOLOv5 algorithm and B/S architecture is designed and implemented.According to the performance comparison results of the two improved algorithms,the improved YOLOv5 algorithm with better performance is adopted.Based on the video monitoring data of the port area,the dynamic visual intelligent supervision is realized for the flow statistics of fishing vessels,the movement of ships in and out of the port,the special control period（typhoon,fishing ban period）and the behavior at night. |
PDF Full Text Request