A adaptive color space merged feature enhancing framework for underwater image enhancement

In response to the core challenges of single color space modeling and low feature fusion efficiency in underwater image enhancement, this paper proposes an adaptive instance feature enhancement bottleneck framework named AIF-FBE. This framework integrates the collaborative learning of Red Green Blue (RGB) and CIELab dual branches. The AIF-FBE uses a dual-branch structure to parallelly extract the intuitive color features (RGB branch) and structural semantic features (Lab branch) of underwater images. Basic blurred features are extracted using ResBlock, and the AIF module is used to achieve adaptive alignment and deep fusion of cross-space features. The AIF module utilizes the AdaIN technique to calibrate the feature distribution differences of the dual branches and combines lightweight convolution and residual connections for precise fusion. The FBE module adopts a "multi-scale convolution encoding attention decoding" structure to restore the blurred features. The encoding end enhances feature interaction and suppresses noise through ConvBlock, and the decoding end improves the spatial resolution through upsampling, the Efficient Local Attention (ELA) mechanism, and deconvolution operations. This method enhances the structural texture details in the blurred areas. AIF-FBE achieved scores of 0.629 and 1.444 on the underwater color image quality evaluation (UCIQE) and non-reference underwater image quality measure (UIQM) quantitative metrics respectively, significantly outperforming existing algorithms. Visually, it can also present more natural seabed textures, clearer object boundaries, and better color balance. This method can provide key technical support for downstream underwater vision applications in practical scenarios.