Enhancing Radar PPI-Based Maritime Object Detection Using Hybrid Preprocessing and Deep Learning Approach

Abstract

In this work, we present a hybrid framework for object detection in maritime radar PPI frames. The method combines domain-specific preprocessing with deep learning models, focusing on YOLOv8s and YOLOv12s. Challenges such as low signal-to-noise ratios, sea clutter, and small target visibility are addressed using adaptive binarization, morphological operations, and noise suppression. A high-resolution P2 feature layer and an area-aware loss function improve small object detection, especially for low-contrast targets such as buoys. The proposed approach is evaluated on 350 simulated radar frames containing annotated fixed objects, moving targets, and landmasses, using five YOLO models (YOLOv5s, YOLOv8s, YOLOv10s, YOLOv11s, YOLOv12s). YOLOv12s achieved the best performance for fixed-object detection on the preprocessed dataset, achieving an mAP50 of 0.772 and an mAP<inline-formula> <tex-math notation="LaTeX">${}_{50-95}$ </tex-math></inline-formula> of 0.503. For the overall evaluation, YOLOv12s obtained the highest mAP50 score of 0.917, while YOLOv8s achieved the best mAP<inline-formula> <tex-math notation="LaTeX">${}_{50-95}$ </tex-math></inline-formula> score of 0.762. After integrating the P2 layer and area-aware loss function, both models produced very similar results. YOLOv8s achieved an mAP50 of 0.957 and mAP<inline-formula> <tex-math notation="LaTeX">${}_{50-95}$ </tex-math></inline-formula> of 0.813, while YOLOv12s achieved an mAP50 of 0.958 and an mAP<inline-formula> <tex-math notation="LaTeX">${}_{50-95}$ </tex-math></inline-formula> of 0.812. In addition, external validation on the real-world DAAN marine radar dataset demonstrated that the gains achieved on simulated data successfully transferred to real radar imagery. The full configuration (preprocessing + P2 layer + area-aware loss) achieved the best performance with mAP50 of 0.922 and mAP<inline-formula> <tex-math notation="LaTeX">${}_{50-95}$ </tex-math></inline-formula> of 0.745, corresponding to substantial improvements over the raw baseline with mAP50 of 0.698, and mAP<inline-formula> <tex-math notation="LaTeX">${}_{50-95}$ </tex-math></inline-formula> of 0.675. These findings indicate promising improvements in the detection of small and low-contrast objects, contributing to reliable real-time situational awareness for autonomous maritime navigation.