Low-light image enhancement method based on hybrid frequency domain-space domain double-branch Mamba

Abstract

The application discloses a low-illumination image enhancement method based on a hybrid frequency domain-space domain double-branch Mamba, relates to the technical field of image processing and computer vision, and comprises the following steps: constructing an image decomposer based on Retinex, generating an illumination hint map, and obtaining a preliminary enhanced image; and constructing a double-branch Mamba module, extracting local textures by using a convolutional neural network branch, and capturing global illumination distribution by using a Mamba branch; wherein the Mamba branch introduces an attention state space module, integrates the illumination hint map as a hint information into a state equation, and supplements context information at a low calculation cost; features are converted to a frequency domain through a Fourier processing module, amplitude and phase components are enhanced respectively, and multi-domain feature fusion is performed in combination with a parallel spatial learning branch; and training is performed based on a hybrid loss function. While reducing calculation overhead, the application effectively realizes collaborative optimization of global illumination recovery and local detail reconstruction, and significantly improves the enhancement quality of low-illumination images.

Images