發明
中華民國
106105547
I 592897
影像辨識加速器系統
國立臺灣師範大學
2017/07/21
一種影像辨識加速系統,其包括: 一影像輸入模組,用以輸入一影像資料;一影像金字塔建構模組,與該影像輸入模組耦接,係預先以軟體找出複數個不同尺度之高斯模板遮罩參數,再透過複數個高斯濾波器模組平行進行複數個卷積運算,其中各所述卷積運算係依該影像資料與一所述遮罩參數進行,以獲得複數個高斯影像,之後,再將所述複數個高斯影像兩兩相減後,輸出至一差分影像模組;一SIFT偵測模組,與該影像金字塔建構模組耦接,係對該差分影像模組輸出之影像資料經由一極值偵測模組及一不穩定特徵點偵測模組進行一極值偵測及一不穩定特徵點偵測運算,以判斷是否為穩定之特徵點,並將該極值偵測、該不穩定特徵點偵測之結果進行一及運算,並儲存至一先入先出暫存器;以及一SIFT描述模組,與該影像金字塔建構模組耦接,係用以對該等高斯濾波器模組輸出之該些高斯影像經由一一階偏微分矩陣模組以及一CORDIC模組進行運算,以求出所有影像點之梯度資料,再以一影像梯度直方圖統計模組及一正規化運算模組對該梯度資料進行運算,以求出該特徵點之描述子資料後,並與該特徵點之位置資料進行結合,俾以提供一即時影像辨識功能。 An SOPC framework based on software and hardware co-design to accelerate the execution of scale-invariant feature transform (SIFT) is proposed, including image pyramid and image feature detection. Because the hardware design is based on pipeline architecture, the execution speed is significantly improved. It is worth mentioning that Gaussian smoothing of image requires exponential function that it is hard to implement and costs a lot of logic elements in hardware. With the use of offline calculation of Gaussian kernel, we can reduce a lot of logic units required in hardware and speed up the system frequency. To eliminate low contrast points, there are quite a lot of inverse matrix operations required in hardware, which results in low performance because many divisors are needed for calculation. To solve this problem, a mathematical derivation is proposed so that use of the divisors can be avoided. As a result, hardware frequency is significantly improved for about twenty times.
產學合作組
77341329
版權所有 © 國家科學及技術委員會 National Science and Technology Council All Rights Reserved.
建議使用IE 11或以上版本瀏覽器,最佳瀏覽解析度為1024x768以上|政府網站資料開放宣告
主辦單位:國家科學及技術委員會 執行單位:台灣經濟研究院 網站維護:台灣經濟研究院