利用圖像特性分類於GPU實現加速單一影像超解析度演算法

A Novel Algorithm of GPU Acceleration for “Super Resolution from a Single Image” Using Patch Characteristic Hashing

邱仲哲(Chung-Che Chiu)

指導教授 ： 曹恆偉

國立臺灣大學/電機資訊學院/電子工程學研究所/碩士(2012年)

https://doi.org/10.6342/NTU.2012.02570

摘要

摘要
超解析度影像演算法可以把相機記錄下的低解析度影像重建出高解析度的影像，重建的影像會帶有高頻的成分，因此會有更多的細節及更清晰的影像。 超解析度的影像可以應用於兩大領域： ﹝一﹞提升人眼觀測到的圖像品質、﹝二﹞幫助機器自動辨識圖形。
由Glasner提出的「單一影像超解析度演算法」是目前眾多超解析度影像演算法中最有潛力的一種，它不需要額外的資料庫，所計算出的超解析度影像也有不錯的品質，但是這種演算法的運算量很大導致實際運用上的困難。 運算量很大的主要原因是此演算法的核心需要使用到「k位最近鄰居演算法」。 我們針對這個問題設計了「圖像特性分類法」以降低運算量，並且使用平行計算晶片來加速「k位最近鄰居演算法」的計算時間。此系統建立於MATLAB的環境下，並用CUDA C撰寫「k位最近鄰居法」加速的部分。並且我們使用「柏克萊影像分割數據庫」來作為測試比較的基準。
實驗的結果顯示我們所提出的方法最終可以使「單一影像超解析度演算法」達到150倍加速。根據測試結果統計，輸出的超解度影像在PSNR方面只會略微下降0.038db，而結構相似指標(SSIM)則只會下降0.009分。經過實驗證明我們所提出的「圖像特性分類法」配合平行計算晶片加速確實可以加快「單一影像超解析度演算法」，同時可以保證輸出的品質維持不變。我們所提出的改良架構不僅試用於「單一影像超解析度演算法」，另外也可以加速相關的碎形超解析度演算法。所提出的「圖像特性分類法」也有從硬體實現的角度加以設計，輔以定點數模擬來保證系統的一致性。我們提出的系統架構十分地具有潛力，加以發展未來可以使超解析度影像演算法實際應用在真實世界。

關鍵字

超解析度 ； 圖形加速卡 ； K-NN搜尋 ； 圖像特性分類法

並列摘要

Abstract
Super resolution imaging is the technology of reconstructing high resolution images with high frequency details from low resolution images recorded by cameras. The needs for high image resolution stem from two application areas: (1) improvement of pictorial information for interpretation; (2) helping representation for automatic machine perception.
“Super Resolution from a Single Image” proposed by Glasner is the most promising method among various super resolution approaches, but its computation time is very long due to high dimensional k-nearest-neighbor search. We proposed a novel patch characteristic hashing method with GPU accelerating k-nearest neighbor search to speed-up the process. Our system is implemented on MATLAB, and we use CUDA C to implement KNN search. The proposed architecture is tested with Berkeley Segmentation Dataset and Benchmark.
The results show that our method can speed-up “Super Resolution from a Single Image” by 150 times faster. The average PSNR is only 0.038dB lower and Structural Similarity (SSIM) only drops by 0.009. The results implicate that the proposed patch characteristic hashing (PCH) can accelerate “Super Resolution from a Single Image” without affecting output quality of the reconstructed images.

並列關鍵字

Super Resolution ； GPU ； K-NN search ； patch characteristic hashing

參考文獻

[4] S. Baker and T. Kanade, "Limits on super-resolution and how to break them," Pattern Analysis and Machine Intelligence, IEEE Transactions on, vol. 24, pp. 1167-1183, 2002.

[6] Y. Min-Chun, H. De-An, T. Chih-Yun, and Y. C. F. Wang, "Self-Learning of Edge-Preserving Single Image Super-Resolution via Contourlet Transform," in Multimedia and Expo (ICME), 2012 IEEE International Conference on, 2012, pp. 574-579.

[7] S. Farsiu, M. D. Robinson, M. Elad, and P. Milanfar, "Fast and robust multiframe super resolution," Image Processing, IEEE Transactions on, vol. 13, pp. 1327-1344, 2004.

[9] H. Ozdemir and B. Sankur, "Assessment of single-frame resolution enhancement algorithms," in Signal Processing and Communications Applications Conference, 2009. SIU 2009. IEEE 17th, 2009, pp. 145-148.

[11] P. Sung Cheol, P. Min Kyu, and K. Moon Gi, "Super-resolution image reconstruction: a technical overview," Signal Processing Magazine, IEEE, vol. 20, pp. 21-36, 2003.