FCA-Net: Accelerating stereo image compression through cascade alignment of side information

Multi-view signal compression, particularly Stereo Image Compression (SIC), plays a pivotal role in applications such as car-mounted cameras and 3D-related scenarios. Despite the Distributed Source Coding (DSC) theory suggesting efficient compression through independent encoding and joint decoding, recent approaches have overlooked the unique characteristics of stereo-imaging tasks, leading to high decoding latency. To address this limitation, we introduce the Feature-based Cascade Alignment network (FCA-Net) to fully exploit side information to accelerate decoding. Initially, we design a feature domain patch-matching module, leveraging stereo priors, reduces redundancy in the search space and minimizes noise introduction. In the subsequent stage, we adopt an hourglass-based sparse stereo refinement network to align inter-image features with reduced computational cost. Experimental results on InStereo2K, KITTI, and Cityscapes datasets demonstrate the superiority of our approach over existing SIC methods. Notably, our approach achieves a decoding speed of 5.67 times faster than the latest DSC-based method, showcasing its efficiency in real-world applications.