Infrared image nonuniformity correction via dual-stream attention and hybrid domain convolution

Abstract

Infrared imaging is widely used in various fields but is often degraded by nonuniformity noise, which poses significant challenges to image quality. Existing nonuniformity correction (NUC) methods often lack accurate modeling of real-world infrared characteristics and struggle to adapt to complex environments. Moreover, many deep learning-based methods originate from visible image processing and are ineffective in addressing the stripe nonuniformity while also exhibiting limited capacity for global feature extraction. To address these issues, we propose a novel infrared image NUC method that integrates dual-stream attention with hybrid domain convolution. A cross-aware attention module is introduced to enhance sensitivity to nonuniformity features such as stripe noise. Combined with a multi-head self-attention mechanism, it forms a dual-stream attention structure that improves global and structural feature modeling. Additionally, we design a hybrid domain convolution module that jointly leverages spatial and frequency information, enabling effective extraction of both local details and global patterns. We also present a realistic simulation method for generating nonuniformity noise in infrared images, facilitating the construction of a high-quality paired dataset for model training and evaluation. Experimental results demonstrate that the proposed method outperforms advanced methods in both visual quality and quantitative metrics, effectively suppressing various types of nonuniformity noise.