LoRA-Augmented ConvMixed-ViT Architecture for Adaptive Compressive Sensing in Resource-Constrained AIoT Scenarios

With the rapid development of Artificial Intelligence of Things (AIoT), the volume of generated data has surged exponentially. Deep compressive sensing (DCS) technology enables accurate data reconstruction at sub-Nyquist sampling rates, significantly enhancing data transmission efficiency and optimizing performance in unimodal vision tasks. However, given the substantial memory and computational resources required, deploying these machine learning (ML) models on resource-constrained edge devices and executing dynamic training poses significant challenges. Furthermore, cloud-based processing introduces privacy vulnerabilities and potential performance degradation. This article introduces a ConvMixed-ViT architecture based on Low-Rank Adaptation (LoRA) and employs an end-to-end methodology integrating learnable CS to facilitate adaptive compression and precise reconstruction on edge environments. Specifically, low-rank adaptive frozen pretrained model weights are injected into the ConvMixed-ViT Architecture’s Transformer variant framework through trainable rank decomposition matrices, greatly reducing the number of trainable parameters for downstream tasks. The experimental results demonstrate that our proposed architecture achieves excellent reconstruction performance on standard benchmark datasets. In common resource-constrained scenarios, the model training scale is not only significantly reduced compared to other algorithms, while the performance is continuously optimized. This demonstrates the superior adaptability and learning efficiency of the model, in dynamic edge computing environments, providing strong support for future edge AI applications in various scenes, including driverless vehicles, smart cities, industrial manufacturing, healthcare services and so on.