特徵擷取是影像分析和辨識中非常重要的一環，越來越多的應用需求朝向即時取得特徵，自組織對映神經網路以及局部二值型態雖然都具有很強的特徵描述能力，但是在軟體執行上的速度無法達到即時需求。因此本研究設計了一個影像色彩和紋理特徵擷取硬體加速器，使用硬體的平行化以及管線化優勢，以加速影像特徵擷取速度。此一硬體加速器具有彈性架構，可提供全域和區域特徵擷取兩個模式，針對SOM色彩特徵量化，我們也設計了2×2以及4×4輸出神經元，提供不同色彩量化硬體架構。實驗驗證對640×480的影像進行特徵擷取，我們的硬體加速器可以在100MHz的系統時脈條件下，相對於PC，可達到10.5倍的色彩特徵擷取速度提升；紋理特徵擷取也比PC快1.875倍的速度，展現了特徵擷取硬體加速器低功耗、體積小且高效率的優勢。;Feature extraction is crucial to image analysis and identification, and applications that need instantaneous feature extraction are increasing. Self-organizing map (SOM) neural networks and local binary patterns possess outstanding feature description capability. However, the computation speed of existent software cannot achieve instantaneous extraction using these two methods. Therefore, this study designed an image color and texture feature extraction hardware accelerator, using the parallelizability and pipelining of hardware to increase the speed of image feature extraction. The proposed accelerator features a flexible framework to enable global and regional feature extraction. To quantize the color features of an SOM, we also designed 2×2 and 4×4 output neurons for different color quantization hardware frameworks. Experiments were performed by extracting features from 640×480 images. Under a system clock rate of 100 MHz, the color feature extraction speed of the proposed accelerator was 10.5 times that of a conventional personal computer. The texture feature extraction speed of the accelerator was also 1.875 times that of the personal computer. The results demonstrated that the proposed hardware accelerator is advantageous for its low energy consumption, small volume, and high efficiency.
