Deep Stereo Image Compression with Decoder Side Information using Wyner Common Information Nitish Mital 1,† , Ezgi Özyılkan 1,† , Ali Garjani 1,‡ , and Deniz Gündüz † † Department of Electrical and Electronics Engineering, Imperial College London ‡ Department of Computer Engineering, Sharif University of Technology {n.mital, ezgi.ozyilkan17,d.gunduz}@imperial.ac.uk, garjania@ce.sharif.edu Abstract We present a novel deep neural network (DNN) architecture for compressing an image when a correlated image is available as side information only at the de- coder. This problem is known as distributed source coding (DSC) in information theory. In particular, we consider a pair of stereo images, which generally have high correlation with each other due to overlapping fields of view, and assume that one image of the pair is to be compressed and transmitted, while the other image is available only at the decoder. In the proposed architecture, the encoder maps the input image to a latent space, quantizes the latent representation, and compresses it using entropy coding. The decoder is trained to extract the Wyner’s common in- formation between the input image and the correlated image from the latter. The received latent representation and the locally generated common information are passed through a decoder network to obtain an enhanced reconstruction of the input image. The common information provides a succinct representation of the relevant information at the receiver. We train and demonstrate the effectiveness of the proposed approach on the KITTI dataset of stereo image pairs. Our results show that the proposed architecture is capable of exploiting the decoder-only side information, and outperforms previous work on stereo image compression with decoder side information. 1 Introduction Data compression is a fundamental and well-studied problem in engineering, and is commonly for- mulated with the goal of designing codes with minimal average code length for a given data ensem- ble. Shannon showed that the entropy is a fundamental bound in lossless data compression when multiple independent samples of the information source can be compressed jointly while allowing arbitrarily small probability of error. The design of entropy codes relies on modeling the probability distribution of the data ensemble. Continuous-valued data (such as vectors of image pixel intensi- ties) must also be quantized to a finite set of discrete values, which introduces error. In this context, known as the lossy compression problem, one must trade off two competing costs: the entropy of the discretized representation (rate) and the error arising from the quantization (distortion). In the case of lossy compression, the fundamental performance bound is characterized by the informa- tion theoretic rate-distortion curve. Recently, deep neural network (DNN) aided data-driven image compression algorithms have received significant research interest, and achieved impressive perfor- mance results, outperforming classical methods, such as JPEG2000 (Skodras et al. [2001]) and BPG (Bellard). 1 Contributed equally to this work. Preprint. Under review. arXiv:2106.11723v1 [eess.IV] 22 Jun 2021