LAUSR

The rapid progress in edge computing (EC) and 5G wireless communication technology has opened up novel opportunities for intelligent applications driven by deep neural networks (DNNs). In particular, machine vision tasks are widely used in mobile/edge computing scenarios. However, the real-time and dense data transmission involved in vision inference services impose significant communication burdens on wireless networks. Thus, this paper investigates the general vision services strategy with cognitive computing network and proposes a communication-efficient edge inference deployment architecture for vision analysis tasks. In this framework, users dynamically perceive the inference data in local, and then compress and offload them to edge servers to perform inference. Specifically, we present a collaborative optimization model of compression ratio and network bandwidth to generate the reliable compression offloading and resource allocation scheme. For this model, the offloading scheme carefully considers the constraints imposed by delay and resources and maximizes the success probability of vision inference tasks. To improve the vision inference performance in the edge network, we further propose a flexible data compression algorithm for images or video frames, which can preserve the more important visual information under a fixed compression rate to reduce the inference accuracy loss from compression. This algorithm first perceives the importance of visual information at different pixel positions, and then compresses different visual regions to varying degrees according to their importance, enabling content-aware adaptive vision data coding. Experimental results show that our proposed offloading model and compression strategy outperform other algorithms, achieving significant communication improvements and performance gains.