This paper introduces a novel deep unrolling architecture that approximates sparse unmixing solutions for hyperspectral image analysis, enabling real-time processing on edge devices. The approach unrolls iterative optimization into a differentiable computational graph, offering end-to-end training with convergence guarantees. Three model variants cater to different deployment scenarios, with the edge-optimized model achieving sub-millisecond latency.
Key findings
Proposes a deep unrolling architecture for sparse unmixing with provable convergence.
Designs three variants for different deployment scenarios, including an edge-optimized model.
Demonstrates 10–100× faster inference compared to classical methods with minimal accuracy loss.
Limitations & open questions
The paper does not discuss the scalability of the model to larger datasets or higher resolution images.
The edge-optimized model's performance under varying network conditions on edge devices is not evaluated.