Research On Ship Target Detection And Recognition Algorithm Based On Deep Learning

Ship detection and recognition technology has significant applications in both civilian and military contexts and has been widely implemented in areas such as maritime traffic management,port management,sea rescue,naval warfare,and intelligence gathering.In the context of port operations,ship targets are often mixed with various interfering objects,resulting in recognition algorithms mistakenly identifying interference as ships or misclassifying ships as other categories.Additionally,the sizes of various ship targets vary greatly,with some small targets being treated as noise or background interference,leading to missed detections in the recognition algorithm.These challenges have brought about severe challenges to target detection and recognition methods,yet current mainstream detection technologies still have some shortcomings in multi-target ship recognition.For example,the YOLO series of algorithms have relatively low accuracy,while the Faster RCNN algorithm has a slower running speed.In order to tackle these challenges,this study focuses on various ships in ports and constructs a ship recognition model based on a pyramid framework to improve the accuracy and recall of the network model.This thesis uses deep learning to investigate ship detection and recognition in the following areas:1.To tackle the problem of small ship targets with fewer pixels and imbalanced ship data categories,we propose a feature pyramid algorithm based on feature decoupling(DFPN).By setting classification scores and location scores as the evaluation metrics for ship identification and location tasks,respectively,and by increasing the proportion of the additive NMS score in the objective function,we increase the network’s ability to recognize fine-grained features.In addition,we introduce a weight factor in the cross-entropy loss function(CE)to overcome the problem of background information overwhelming ship target information and the imbalance between large and small ships.Our experiments show that DFPN achieves better recognition accuracy in identifying small ship targets compared to YOLOv4 and Faster R-CNN,with an improvement of 0.2% and 1.5% in the m AP performance index,respectively.2.To address the issue of multi-scale ship identification and improve the speed of the detection algorithm,we propose a feature pyramid algorithm(FDFPN)that incorporates an attention mechanism.Firstly,we use the residual edge of the gated attention mechanism to eliminate some useless parameters and improve network identification efficiency.Then,we design a multi-feature fusion module based on ship features that can fuse multipledimensional ship features to improve the recall rate of target detection.In comparative experiments,FDFPN achieves an F1 score of 77.5%,an m AP score of 80.9%,and a FPS of41.Compared to YOLOv4 and Faster R-CNN,FDFPN improves the m AP performance index by 6.3% and 7.6%,respectively.Experimental results demonstrate that FDFPN has higher accuracy and recall rate in multi-target ship identification.