OIPF: An Orthogonal Inputs Perception Fusion Framework for Infrared Small Target Detection

In infrared small target detection (ISTD), a common challenge for neural network methods is that as the network deepens, the sparse information of small targets becomes even more diffuse in the deeper layers, limiting the ability to extract high-level semantic features. To address this issue, we propose the Orthogonal Inputs Perception Fusion (OIPF) framework. The key idea is to enhance the learnable target features in deeper layers without increasing network depth. This is achieved by creating orthogonal data pairs through intensity inversion and enhancement, which are then fed into a dual-input framework to provide incremental information flow. Additionally, we introduce a Relationally Aware Module (RAM) that generates spatial weight maps by leveraging the relationships between data pairs across layers. This module helps the framework focus on target edges and complex background regions, ensuring that rich target information is maintained throughout the network. Through extensive testing on three datasets and ablation study, we validated the OIPF framework and RAM module's superiority, as well as their low dependency on dataset scale. By integrating these into existing models, we significantly enhance ISTD performance, proving our solution's effectiveness and robustness.