Dynamic image compression and reconstruction via tensor decomposition on edge nodes

In wireless multimedia sensor networks, edge nodes are constrained by computational resources and energy supply, necessitating an efficient balance between image compression and reconstruction. To address this, this paper employs block-sparse tensor-based compression coding for images. A virtual codebook pool with block-sparse characteristics is trained based on image texture features, utilising Tucker decomposition and fractal coding for greyscale matching. Building upon this, a hierarchical clustered network topology is designed to collaboratively perform image decomposition, compression, and reconstruction at edge nodes. To enhance image reconstruction quality, a dynamic image reconstruction model based on block-sparse tensor decomposition and the transformer architecture is proposed. Block-sparse tensor decomposition is embedded within the transformer to learn global information of the image. Experimental results demonstrate that the proposed method achieves a network energy consumption of only 397.48 nJ/bit, with a peak signal-to-noise ratio of 38.86 dB.