<p>Binocular stereo images are employed in many imaging systems to create a 3D effect. To make easy the access and exploitation of stereo image databases, content based retrieval solutions are required. This paper proposes novel deep learning based architectures to deal with content based retrieval specifically dedicated to color stereo images. The proposed method consists of three main parts. First, a preprocessing step is performed. It is based on a discrete wavelet transform (DWT) to generate different complementary scale-space representations of each view and the depth map. Then, the feature extraction step consists in applying neural networks to each obtained DWT subband to obtain their related deep features. Finally, in the fusion step, these features are aggregated to produce a global feature representation of the stereo images. Unlike previous deep learning-based stereo image retrieval methods, we investigate deep fusion schemes that leverage attention mechanism. The experimental results show the effectiveness of the proposed approach compared to the state-of-the-art methods.</p>