FREGNet: Ship Recognition Based on Feature Representation Enhancement and GCN Combiner in Complex Environment

Abstract

Harsh sea conditions, uneven illumination, and the variable spatial positions of ships result in ship images captured by imaging systems that contain not only the ship targets but also complex environmental information. This environmental complexity, intertwined with the ship targets, significantly undermines the accuracy of ship target recognition. However, existing methods for recognizing ships are mainly designed to identify large targets in clear weather with uniform illumination. They rarely account for the aforementioned complex environmental information, leading to the loss of critical detailed features in the ship’s target area, particularly in adverse weather and uneven illumination scenarios. To address these challenges, we propose a novel method for recognizing ship targets in complex environments: the Feature Representation Enhancement and Graph Convolutional Combiner Network (FREGNet). We introduce a Feature Representation Enhancement (FRE) module to the backbone of the feature extraction network to enhance the capture of detailed features in the ship’s target area, particularly under illumination noise and cloud-covered conditions. Furthermore, we design a GCN Combiner module based on a Graph Convolutional Network to dynamically combine global feature key points and local detailed feature key points output by the FRE, thereby increasing the quantity of detailed feature information in the ship’s target area. This approach facilitates the accurate classification of ship categories. Experiments were conducted using the CIB-ships, MAR-ships, and Game-of-ships datasets. Compared to suboptimal transformer-based ship target recognition methods, FREGNet improves ship target recognition accuracy by 2.8%, 0.99%, and 1.38%, respectively. The FREGNet method achieves high recognition accuracy at a faster rate.