Deep learning has become a cornerstone of modern stereo matching algorithms due to its ability to accurately model scene geometry. Disparity estimation varies depending on the specific constraints of each application: some methods rely on recurrent architectures to dynamically refine predictions by focusing on uncertain regions, while others exploit multi-scale strategies, processing images at different resolutions to capture fine details in visually complex environments.

In this context, we propose WS-CRNN, a hybrid architecture that combines the reactivity of recurrent refinement with the structural richness of multi-scale analysis. At the core of the model, the wavelet scattering transform robustly extracts both local and global features at multiple scales, while substantially reducing the dimensionality of cost volumes, thus alleviating the computational overhead typically associated with 3D convolutions.

In parallel, a recurrent neural network iteratively enhances disparity predictions through successive comparisons of left and right views, enabling precise hierarchical estimation. Compared to state-of-the-art methods, WS-CRNN achieves competitive performance while maintaining low memory and energy consumption. Overall, WS-CRNN represents a promising trade-off between algorithmic complexity and prediction quality, marking a significant step forward in the field of deep stereo vision.

Index Terms- Deep learning, Stereo matching, Wavelet scattering, CNN, RNN, ConvLSTM, Multi-scale, Matching cost volume, Disparity map.