A deep learning denoising model method and system based on frequency domain enhancement and entropy-guided scanning

Abstract

This invention discloses a deep learning denoising model method and system based on frequency domain enhancement and entropy-guided scanning, comprising the following steps: multi-scale frequency domain decomposition of input data, and fusion of features from different frequency bands using an attention mechanism to generate frequency domain enhanced features; traversing the input data through a multi-scale sliding window, calculating the complexity distribution of local regions within each window, generating a multi-scale entropy map, and adaptively generating a scanning path that prioritizes coverage of high-complexity regions based on the entropy map content; inputting the feature sequence corresponding to the scanning path into the state space enhancement module, and sequentially performing residual fusion of spatial and spectral features, serialized feature extraction, long-range dependency modeling, and bidirectional gated fusion to obtain the enhanced deep feature representation; spatially aligning and channel-fusioning the frequency domain enhanced features with the multi-scale features output by multiple state space enhancement modules, and reconstructing the denoised data by the decoder; this invention has good versatility and scalability.

Images