Refined Anchor-Free Model With Feature Enhancement Mechanism for Ship Detection in Infrared Images

Ship detection in infrared images has found wide applications in both civil and military fields. Nowadays, infrared ship images serve as an irreplaceable information source for marine surveillance tasks due to the all-day and all-weather imaging capabilities of infrared sensors. However, previous researches in optical and synthetic aperture radar (SAR) images can not be simply applied to ship detection in infrared images, there are multi-scale ship targets in infrared datasets, especially small-sized targets with dense distributions. Besides, the complex surroundings including inshore buildings and thick clouds put higher difficulties for ship detection tasks. In order to handle the above problems, we propose a refined anchor-free model with feature enhancement mechanism for ship detection in infrared images. Firstly, we redesign the network of YOLOv8 as the anchor-free detection model, aiming to perform the best multi- scale detection accuracy and running speed. Secondly, a multi- attention fusion (MAF) module based on convolution kernels with different sizes is embedded into the backbone to distinguish the spatial salience of the targets from the complex surroundings. Besides, for improving precision in multi-scale ship detection task, we propose the refined feature alignment (RFA) and label assignment redesign (LAR) modules, the former is used during the feature fusion process to cope with the problem of contextual misalignment, and the latter is introduced to refrain from the disadvantageous effect of the fuzzy areas in the process of model training. Experiments are conducted to evaluate the detection accuracy of the proposed model on the public infrared datasets ISDD which prove the model's superior than the state-of-the-art detection models for ship detection task in infrared images