Ship Detection Algorithm Based on Structural Reparameterize Dilated Large-Kernel Convolution and Spatial Selective Kernel Mechanism

In response to the issues of large target scale variations and complex background environments in ship detection, this paper presents a single-stage ship detection algorithm named DrbLSK, which is an improvement on YOLOv5. Firstly, to enhance the model's ability to model the context of targets with different scales, a kernel selection module is introduced into the backbone network. By dynamically selecting the receptive field, environmental interference is reduced. Simultaneously, the combination of large-kernel dilated convolution and structural reparameterization is employed to reduce model parameters and enhance the feature expression capability of the backbone network. Next, a decoupled detection head is utilized to alleviate the conflict between classification and regression tasks in the target detection task. Moreover, CIoU is used in the detection head to replace the original loss function, thereby accelerating the convergence of the network. Experimental results show that, on the ABOships dataset, the improved model reduces the number of parameters while increasing accuracy by 2.5\% compared to YOLOv5, with a mean average precision of 63.9.