An Improved Convolutional Layer Based on Stochastic Masked Kernel for Ship Target Detection

Ship target detection is crucial for ensuring maritime safety. To tackle the issues of limited recognition accuracy and inadequate generalization in existing detection models, this thesis presents an enhanced model: YOLOv5s-Mask, which incorporates an Improved Convolutional Layer based on Stochastic Masked Kernels. This novel approach aims to boost both recognition accuracy and the model's ability to generalize across various scenes. We introduce the concept of the Stochastic Masked Kernel and develop the Masked SPPF (Masked Spatial Pyramid Pooling) module to improve the model's detection performance and robustness, particularly for small and densely packed targets. Additionally, we incorporate the SE (Squeeze-and-Excitation) attention mechanism to further refine recognition accuracy while keeping the model lightweight. Experimental results show that YOLOv5s-Mask achieves a 1.26% improvement in mAP and a 5.71 % increase in mAP0.5-0.95 compared to the original YOLOv5s. This demonstrates the model's significant potential for real-world ship target detection applications.