| Sea surface target detection has always been of great significance in marine cognition and sea surface surveillance.Sea surface target detection is mainly divided into two major problems:The first is sea clutter suppression under high sea conditions and non-weak targets;The second is the problem of target detection of low sea conditions and weak targets.First of all,sea clutter suppression has always been a difficult point in the field of sea surface target detection,and traditional sea clutter suppression algorithms are difficult to model complex sea clutter data well.However,when RNNs such as Long Short Term Memory(LSTM)network are used to predict and suppress sea clutter based on the temporal correlation of sea clutter,there are problems such as gradient diffusion and slow training speed caused by long input sequences.Secondly,the focus in low sea conditions is on the detection of small targets.Problems such as low signal-to-clutter ratio of weak targets will make it difficult for radar to detect targets.Finally,with the improvement of real-time requirements for sea surface target detection,the requirements for algorithm detection speed are also getting higher and higher.To solve these problems,a sea clutter suppression algorithm based on Temporal Convolutional Network(TCN)prediction model,an improved YOLOv5 sea surface small target detection algorithm with combined multi-frame,and a lightweight improved YOLOv5 model are designed.The research contents of this paper are as follows:1.The sea clutter suppression algorithm based on TCN prediction model was studied.First,the TCN sea clutter normalized amplitude sequence prediction model is established to predict sea clutter,and then a sea clutter suppression algorithm based on the time domain cancellation after TCN prediction is proposed,and finally,based on the measured data of sea radar,the proposed model algorithm is compared and analyzed with the existing model.The experiments show that the proposed algorithm can effectively predict and suppress sea clutter.2.The multiframe detection approach for sea surface small target by using improved YOLOv5 was studied.A multi-frame detector for small sea surface targets based on the improved YOLOv5 implementation is proposed to improve the detection accuracy and reduce false alarms and missed alarms.Firstly,the structure of YOLOv5 network is improved:Spatial Pyramid Pooling(SPP)of the original network is replaced with Receptive Field Block(RFB)module.Feature Pyramid Network(FPN)and Path Aggregation Network(PAN)in the multi-scale fusion structure were replaced by Bidirectional Feature Pyramid Network(BiFPN).The Simple,Parameter-Free Attention Module(SimAM)has been added to pay more attention to the characteristics of small targets.Secondly,the original radar echo image is fused with its Constant False Alarm Rate(CFAR)processing results and sent to the network for detection.Finally,the original backtracking filter bank is improved for multi-frame accumulation detection.This method has verified the measured data of the radar scanning mode.The detection results show that the algorithm can effectively improve the detection performance of small targets on the sea surface.3.The rapid detection of sea surface targets by the lightweight improved YOLOv5 was studied.Firstly,some lightweight neural networks are used to replace the original backbone network BottleneckCSP in the improved YOLOv5 model to achieve the lightweight of the sea surface target detection model.After experimental performance comparison,ShuffleNet V2 was finally selected to replace the original backbone network.After the initial realization of the lightweight task,the standard convolution of the detection head part is replaced with a depthwise separable convolution to further lightweight the model.The final experimental results prove that the proposed method can make the overall detection model lightweight and realize the rapid detection of sea surface targets without losing more detection accuracy. |
PDF Full Text Request