Underwater image cross-channel attention frequency domain decoupling enhancement network, system and method
Patent Information
- Application Number
- CN202610933234.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-26
- Publication Date
- 2026-10-16
AI Technical Summary
[0008]本发明的目的是提供水下图像跨通道注意力频域解耦增强网络、系统及方法，旨在解决现有水下图像增强中颜色通道退化差异建模不足、低频色彩恢复与高频纹理保护相互干扰，导致色彩补偿不充分和细节易受破坏的技术问题
[0036] This invention maps degraded underwater images into initial channel features through an input projection layer, and decomposes these initial features into low-frequency and high-frequency sub-bands using a Haar frequency domain decoupling enhancement module. The low-frequency sub-band primarily carries global brightness and color distribution information, while the high-frequency sub-band primarily carries edge, texture, and local detail information. For the low-frequency sub-band, this invention uses a cross-channel attention correction module to generate adaptive fusion weights for the red, green, and blue color channels based on global statistical information, constructing cross-channel reference features. These cross-channel reference features serve as a unified source of key and value features, and cross-attention correction is applied to each color channel separately, thus establishing inter-channel compensation relationships to address the issue of inconsistent attenuation levels among different color channels in underwater imaging. For the high-frequency sub-band, this invention enhances edges and texture details through a local texture enhancement path, reducing interference from the color correction process on high-frequency details. Subsequently, frequency domain enhancement features are reconstructed using Haar inverse discrete wavelet transform, and spatial context relationships and local texture continuity are further enhanced through a spatial refinement module. Finally, an RGB enhanced image is output from the image reconstruction branch. Simultaneously, during the training phase, the physical constraint branch predicts the transmittance map and background scattered light based on spatially refined features or the RGB enhanced image, and reconstructs the degraded image based on the underwater image degradation model. This ensures consistency between the reconstructed degraded image and the input degraded underwater image, constraining network training. Therefore, this invention can improve problems such as color shift, channel imbalance, low contrast, and blurred details in underwater images, enhancing the color preservation ability, structural clarity, and physical consistency of the training process in the enhanced image.