基于DHFC-DETR的近海船舶检测算法OA

To achieve efficient and accurate detection of nearshore ship images,and address challenges such as heavy background noise interference,severe object overlap,and low recognition accuracy in nearshore ship detection,this paper proposes a DETR detection model utilizing dynamic hypergraph fusion computation,and the model is based on improved RT-DETR.A deformable attention(DA)mechanism is introduced to the attention-based intra-scale feature interaction,which reconstructs regions of interest during attention queries and provides more accurate deformation areas for feature points.Additional feature extraction branches are added to the backbone network,paired with a new neck structure.Here,image features are subjected to hypergraph convolution and progressive stepwise fusion,and the feature information is lifted to higher dimensions.The similarities between feature points are evaluated to aggregate similar ones while filtering out irrelevant ones,and finally the refined feature information is obtained by progressive stepwise fusion.Finally,dynamic scale fusion performs feature extraction and dimensionality reduction compression on the extracted features within a three-dimensional space.Combined with the dynamic concatenation module,this process enables interaction of features across different scales.Moreover,a new complex nearshore ship detection dataset is constructed to verify the object detection performance of the model in complex scenarios.Experiments show that the proposed model surpasses existing ones in mAP on both the self-constructed dataset and the publicly available IEEE SeaShips dataset.To sum up,the effective detection of nearshore ships is realized.