FreqODEs: Frequency Neural ODE Networks for Infrared Small Target Detection

Tianxiang Chen, Zi Ye

Abstract: Infrared small target detection (ISTD) is aimed at segmenting small targets from infrared images and has wide applications in military areas. With the specially designed spatial domain feature extraction modules, recent ISTD networks have already achieved promising performance. However, almost all of them are prone to fail in cases where the small targets are so similar to their surroundings that they look seamlessly integrated into the background. The main reason is that only spatial feature domains are used, which is powerless in this situation. To enhance the ability to distinguish extremely similar cases, this study proposes extracting features from the frequency domain. Specifically, we drive self-attention operation to interact with frequency information and propose frequency interaction attention (FIA). It draws frequency clues from the input feature by applying a wavelet transform to achieve reversible down-sampling before interacting with the spatial domain feature so that small targets with high similarity with the background can be highlighted. However, some high-frequency components, including boundaries and noise, may also be highlighted and interfere with the target areas’ discrimination. To fully suppress these irrelevant disturbances, we devise a family of neural ordinary differential equation (ODE) modules stacked behind FIA based on high-accuracy ODE solvers, where we introduce coefficient learning to block design to lower the risk of gradient vanishing that most neural ODE networks have. Based on these two designs, we build a family of frequency neural ODE networks, namely FreqODEs. The experiments performed on two of the most widely used ISTD datasets, NUAA-SIRST and IRSTD-1k, show the superiority of the proposed method.