MK-IGEV: collaborative integration of Mamba feature extraction and KAN cost aggregation for underwater stereo matching

Current stereo matching methods based on convolutional neural networks (CNNs) and iterative optimization frameworks face substantial challenges in distortion-resilient feature representation and physics-aware modeling of underwater light scattering and attenuation. This study proposes an underwater stereo matching network, MambaKAN-IGEV (MK-IGEV), which integrates Mamba feature extraction and KAN cost aggregation in a collaborative manner. Specifically, the network combines visual state-space modeling with the Kolmogorov-Arnold Network (KAN). In the feature extraction stage, the MambaExtractor module employs a selective scanning mechanism to achieve adaptive spatial association of degraded features, addressing local feature distortion caused by underwater suspended particle noise. In the cost aggregation stage, the proposed method constructs a ResUKAN module based on third-order B-spline basis functions to explicitly model the physical laws of light attenuation. By integrating interpretable constraints through a residual architecture, the module enhances the traceability of model decisions. Ablation studies and comparisons with other state-of-the-art models demonstrate the effectiveness of MK-IGEV. It achieves the best performance across multiple benchmarks with only 15.47 M parameters. Additionally, in engineering application evaluation, the proposed method achieves high-precision short-range underwater measurements with the lowest mean relative error.