Integration of Multidimensional Features for Infrared Image Enhancement

Due to the characteristics and equipment limitations, infrared images often experience a series of degradations, such as noise and low contrast. Deep learning-based image enhancement methods are both efficient and flexible, often involving a progressive reduction in image size to extract crucial features. However, due to the simplicity of infrared image composition, that's not viable. Compressing the resolution may result in the loss of spatial details, making it hard to recover high-quality images through feature extraction. In this context, we propose a new architecture designed to keep the infrared image size consistent throughout the network, thereby preserving spatial accuracy. Additionally, it integrates complementary contextual information from low-resolution representations. For feature integration, we introduce a self-attention mechanism, aggregating two aspects of features: (1)Multi-scale parallel feature streams, which reflect spatial details at different resolutions; (2)Sooner and subsequent feature streams, which reflect contextual details. Despite the limited information capacity of infrared images, our method can capture rich features, achieving multiple image enhancement while maintaining the original details. Our model sets a new standard in enhancing infrared images, proven by tests on various datasets.