- 學位論文
利用拋物曲線攝影進行去模糊計算之硬體加速器設計
Hardware Accelerator Design for Parabolic-Photography-Based Deblurring
摘要
運動模糊是相當常見的影像模糊成因之一，主要為物體在相機的快門時間內移動所導致。若有多物體在場景中運動時，每個物體相對於相機所形成的運動模糊取決於該物體的相對運動速度，不同的運動速度會產生不同的模糊核，使得去模糊變得相當的困難。 在本篇論文中，我們利用相機拍攝時的拋物曲線運動去補償物體的運動，即可得到接近單一模糊核之影像，再配合自然影像的先備知識進行去模糊運算，即可得到清晰影像。此外探討了拋物曲線攝影的先天限制條件，模糊核的估計方法與其去模糊演算法。 由於去模糊演算法在運算時資料交換量龐大，使得傳統運算系統在執行時硬體效率不佳。利用硬體平行化技巧，我們以積體電路實作出拋物曲線去模糊處理器。輸入尺寸為640x480之拋物曲線攝影圖像與模糊核，可於1.1秒內完成去模糊運算，與軟體相比加速倍率可達14倍。使用的製程為TSMC 90 nm、運作頻率為125 MHz、晶片尺寸為7.29 mm2、消耗功率為438.74 mW。
並列摘要
Motion blur is a common cause of image blur. It is mainly induced by object movement during exposure. In a scene with multiple moving objects, different objects project different blur images on the sensor plane depending on its moving speed with respect to the camera. Since different moving speeds would induce different blur kernels, the deblurring process could be very challenging. In this thesis, we investigate the limits of parabolic photography, the effectiveness of the blur kernel and its corresponding deblurring algorithm. The parabolic movement of the camera can compensate the movement of objects. By using this particular movement pattern, we can use the same blur kernel for different objects yet receive good deblurring results. With the help of natural image prior, we can further improve the sharpness of images. Since deblurring algorithm involves huge amount of data exchange, the performance of conventional computing systems would be degraded for low utilization efficiency. With TSMC 90 nm technology, we design a processor which operates at 125 MHz and is capable of deblurring a 640x480 blurred image captured by the parabolic camera within 1.1 second. It is 14 times faster than the software version. The chip area of 7.29 mm2 with the power consumption of 438.74 mW.