SMFR-Net: simple multi-domain flare removal network.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Image flare, caused by light reflections and scattering, degrades image quality and impacts vision tasks.
• Existing Transformer models struggle with computational complexity at high resolutions, while windowed attention limits receptive fields.

Purpose of the Study:

• To develop an efficient and effective network for image flare removal.
• To address the limitations of existing Transformer architectures in handling image flare.

Main Methods:

• Proposed SMFR-Net, a multi-domain network with an encoder modeling frequency and spatial domains, and a simplified decoder.
• Encoder utilizes Fourier Transform for global context and dilated convolutions for expanded receptive fields, incorporating Channel-Spatial Attention.
• Designed a structure-aware composite loss function to enhance performance.

Main Results:

• SMFR-Net achieved state-of-the-art performance with 7.981M parameters.
• Outperformed existing methods on Flare7K++ and synthetic datasets across multiple metrics.
• Demonstrated superior flare removal and potential for real-world applications.

Conclusions:

• SMFR-Net offers an efficient and effective solution for image flare removal.
• The multi-domain approach and simplified decoder contribute to its strong performance and efficiency.
• The network shows significant potential for practical applications in image enhancement.