應用於微型無人機遙測之高光譜影像處理器

Abstract

本文提出一款專用處理器，實現了高光譜影像（HSI）處理中完整的光譜解混流程，涵蓋降階（rank reduction）、端元（endmember）提取以及豐度（abundance）估計三個主要階段。為了在降低複雜度的同時保持高效能，本設計採用了多項創新的硬體優化技術。該架構中的處理單元（processing elements, PEs）採用摺疊（folding）與資料交錯（data interleaving）技術，以降低硬體複雜度並維持運算能力；同時配置深度流水線，以加速解混過程中的高運算量操作。在演算法層面，採用了基於豪斯霍爾德變換的單體成長（HTSG）演算法進行端元提取，並結合 L1 正則化的影像空間重建（ISR）方法進行豐度估計，在計算效率與解混精確度間取得平衡。此外，引入稀疏度自適應時脈控制機制（sparsity-adaptive clocking），利用資料稀疏性以降低動態功耗。 該處理器採用 40 奈米 CMOS 製程實現，核心面積為 2.56 mm²，在 0.68 V 供應電壓與 175 MHz 時脈下運行，功耗為 44.3 mW。此設計可同時生成 8 個端元及其對應的豐度圖，針對尺寸為 256×256×64 的 HSI 影像，運算吞吐率達到 62.4 張/秒。經實測，在 HYDICE Urban 數據集上可達到 33.2 dB 的峰值訊噪比（PSNR），驗證了本設計在實際應用場景中的有效性。 與高階 CPU 相比，本處理器在運算速度上提升 540 倍、能源效率約提升 1,700,000 倍、面積效率約提升 31,000 倍；與高階 GPU 相比，分別提升 17.5 倍、230,000 倍與 4,000 倍。透過創新的硬體架構設計與優化策略，本處理器能在功耗與尺寸限制下，達成即時高光譜解混運算，適用於電池供電的微型飛行載具（MAV）等遙測應用。

This paper presents a dedicated processor that implements the complete spectral unmixing workflow for hyperspectral image (HSI) processing, covering three key stages: rank reduction, endmember extraction, and abundance estimation. To achieve high performance while maintaining low complexity, the design incorporates several hardware optimization techniques. The architecture integrates folding and data interleaving techniques to reduce hardware complexity while maintaining computational capability. The processor employs deep-piped reconfigurable processing elements (PEs) to accelerate compute-intensive operations in the unmixing process. Through algorithm-architecture co-optimization, this processor achieves both high performance and hardware efficiency. It implements the Householder Transformation-based Simplex Growing (HTSG) algorithm with parallel CORDIC arrays for endmember extraction, while utilizing a simplified L1-regularized Image Space Reconstruction (ISR) scheme with folded processing elements for abundance estimation. A sparsity-adaptive clocking mechanism is introduced to exploit data sparsity and reduce dynamic power consumption. The processor is fabricated in 40-nm CMOS technology, with a core area of 2.56 mm². It operates at 175 MHz from a 0.68-V supply and consumes 44.3 mW of power. The design supports the concurrent generation of 8 endmembers and their corresponding abundance maps for a 256×256×64 HSI, achieving a throughput of 62.4 frames per second. When tested on the HYDICE Urban dataset, the processor achieves a peak signal-to-noise ratio (PSNR) of 33.2 dB, validating its effectiveness in practical applications. Compared to a high-end CPU, the processor achieves a 540× increase in processing speed, 1,700,000× improvement in energy efficiency, and 31,000× improvement in area efficiency. Compared to a high-end GPU, it achieves 17.5× higher processing speed, 230,000× higher energy efficiency, and 4,000× higher area efficiency. Through innovative architectural design and optimization strategies, these results demonstrate that the processor can perform real-time hyperspectral unmixing within power and size constraints suitable for onboard deployment, such as in battery-powered micro air vehicles (MAVs) for remote sensing applications.