A Study on Deep Learning-Based All-Day Port Ship Detection Algorithms

To address the potential issues of insufficient detection accuracy, numerous small targets, and diverse ship types in complex port environments, this paper proposes a port ship target detection and recognition method based on an improved YOLOv8 model. Furthermore, to meet the requirements of all-weather ship target detection, the study utilizes datasets from both visible light and infrared modalities. Considering the limited availability of image acquisition equipment in ports, a single-model detection algorithm is developed for handling data from both modalities. Firstly, an Enhanced Feature Correlation (EFC) technique is introduced to replace the simple additive fusion method across different hierarchical levels in YOLOv8's feature fusion mechanism. This modification aims to mitigate the loss of small-target information caused by multiple downsampling operations. Secondly, to address the challenges posed by complex backgrounds and occlusion between ships, a Change Boundary-Aware Module (CBM) is incorporated into the network. By capturing boundary information of dynamic regions, the CBM helps the model to more effectively focus on ship target areas. Finally, to cope with poor target-background distinction in low-contrast scenarios such as nighttime or inadequate lighting conditions, a Parallelized Patch-Aware Attention (PPA) module is introduced. This module enhances the model's ability to suppress noise and handle low contrast, thereby improving the distinguishability of ship targets from the background. Experimental results show that, compared to the original YOLOv8 model, the improved algorithm achieves a 1.9 % increase in detection performance on the visible light dataset and a 0.9 % increase on the infrared dataset.