Edge-Oriented Infrared Ship Pattern Recognition in Complex Maritime Scenes via Deep Feature Enhancement and Teacher-Guided Distillation

Hongliang Tian¹, Chenying Pei^1,*, Jin Lei², Xiaoke Liu¹, Xin Ma³
1 Key Laboratory of Modern Power System Simulation and Control & Renewable Energy Technology, Ministry of Education (Northeast Electric Power University), Jilin, China
2 School of Marine Science and Technology, Northwestern Polytechnical University, Xi’an, China
3 Micro Engineering and Micro Systems Laboratory, School of Mechanical and Aerospace Engineering, Jilin University, Changchun, China
* Corresponding Author: Chenying Pei. Email: email
(This article belongs to the Special Issue: Machine Learning and Deep Learning-Based Pattern Recognition, 2nd Edition)

Computer Modeling in Engineering & Sciences https://doi.org/10.32604/cmes.2026.087611

Received 19 June 2026; Accepted 28 August 2026; Published online 17 September 2026

Download PDF

Abstract

Infrared ship detection is an important deep learning-based pattern recognition task for maritime visual perception, where accurate target recognition under complex thermal backgrounds is essential for intelligent monitoring and real-time decision support. However, low target-background contrast, sea-wave thermal textures, coastline heat-source interference, and specular thermal reflections in infrared maritime imaging weaken discriminative ship patterns and reduce recognition reliability in complex scenes. To address these challenges, we propose an edge-oriented infrared ship detection method for real-time maritime monitoring. The proposed method reconstructs the feature pyramid by integrating a Wavelet-Frequency Enhancement Module (WFEM) with a Dynamic Multi-Scale Feature Fusion Module (DMS-FFM), thereby enhancing thermal clutter suppression, small-target pattern representation, and multi-scale feature interaction in dense maritime environments. A Task-Conditioned Unified Detection Head (TCUDH) is further introduced to improve localization robustness through shared representation learning and task-conditioned feature modulation, while retaining structural extensibility for related visual prediction tasks. In addition, a Teacher-Guided Dual-level Structured Knowledge Distillation (TDSKD) strategy improves student classification and localization through prediction-structure and geometric consistency distillation. Experimental results demonstrate that the proposed method achieves excellent detection performance across multiple datasets while maintaining low computational overhead, with 3.71M parameters, 22.2 giga floating-point operations (GFLOPs), and 234.7 frames per second (FPS) on an NVIDIA GeForce RTX 3090 graphics processing unit (GPU). Moreover, the proposed model achieves an average effective inference throughput of 28.23 FPS on a Jetson Orin Nano 8 GB edge device, demonstrating its potential for real-time edge-side pattern recognition and visual perception in infrared maritime monitoring tasks.