ConvMixed-ViT Architecture Based on LoRA for Compressive Sense in 6G-AIoT at Resource Constrained Environment

With the rapid development of 6G networks and artificial intelligence of things(AIoT), the volume of generated data has surged exponentially. Deep Compressed sensing(DCS) achieve accurate data reconstruction at Sub-Nyquist rates, minimizing data transmission volume and optimizing performance in unimodal vision tasks. However, given the substantial memory and computational resources required, deploying these DCS models on resource-constrained edge devices and executing dynamic training poses formidable challenges. Additionally, cloud-based processing introduces privacy vulnerabilities and potential performance degradation. This paper introduces a ConvMixed-ViT architecture based on Low-Rank Adaptation (LoRA) and employs an end-to-end methodology integrating learnable CS to facilitate compression and precise reconstruction on edge environments. Specifically, the LoRA freezes pre-trained model weights and injects them into the ConvMixed-ViT Architecture's Transformer variant framework through trainable rank decomposition matrices, greatly reducing the number of trainable parameters for downstream tasks and controlling the model size. The experimental results demonstrate that our proposed architecture achieves excellent reconfiguration performance on standard benchmark datasets. In resource-limited environments, CTLCS offers efficient adaptability and optimizes performance, proving its suitability for future edge AI applications.