OFSPNet: ODE-guided fusion network with the same paradigm for infrared small target detection

OFSPNet: ODE-guided fusion network with the same paradigm for infrared small target detection

연구 분야: Software Development

학회: The Journal of Supercomputing

Infrared small target detection (ISTD) remains a challenging task due to the complex backgrounds and indistinct structural features of small targets. Current ISTD methods have the following limitations: on the one hand, both CNN-based and transformer-based methods have deficiencies in feature extraction. On the other hand, current fusion methods cannot effectively utilize global information to enhance the local features of the target, which hampers further improvement in detection performance. This paper proposes the OFSPNet model, which significantly improves detection accuracy and detail-capturing ability through diversified feature representation and adaptive fusion. Firstly, the model employs a parallel architecture integrating CNN, transformer, and Mamba—CNN is adept at extracting low-level local details, while transformers focus on capturing global information. Meanwhile, Mamba relies on its unique selection mechanism to filter out irrelevant data and retain the important features of key visual cues in the image. Secondly, an ODE-inspired adaptive feature fusion module was designed as an information bottleneck to suppress high-frequency noise, while simultaneously strengthening target features through backpropagated gradients. Then, a modular design is carried out for the transformer and Mamba structures. A MLP is added to enhance the ability to handle nonlinear problems, and a gating mechanism is introduced to dynamically adjust information flow, suppressing noise and irrelevant background information. Finally, considering the real-time requirements in practical applications, the model has achieved a balance between inference speed and performance. Through optimization strategies, the inference speed has been increased by 15.28%, ultimately achieving the dual advantages of efficient inference and high detection accuracy.The code are available at https://github.com/1yanchen3/12.git.