近年來，隨著科技技術進步，人類對於影像上的應用需求，從傳統的二維平面影像，逐漸轉換成三維立體空間的影像；亦即在影像資料當中，多了『深度』信息的需求，深度訊息可以被廣泛運用於各類的前瞻系統，例如: 自動駕駛技術、擴增實境 (AR)、虛擬實境 (VR)、機器人導航判斷…等。因此，如何設計有效的深度預測演算法以產生高精確度的深度資訊是一項重要且急迫的課題。
在各項深度預測演算法當中，最簡單且有效的方法為採用雙鏡頭的深度法。此方法首先藉由兩個平行擺放的鏡頭拍攝圖像，並設計一套視差預測演算法以確立圖像之間的視差值；接著配合鏡頭焦距、鏡頭間距離等已知資訊，產生出最終深度結果。因此，視差預測演算法可視為雙鏡頭深度法當中的關鍵步驟，其錯誤率對於最終的結果具有極大的影響。近年來為了效能上的提昇，許多研究也將其所提出的視差預測演算法以數位電路進行實現。然而，在導入實際應用的過程當中，仍然有需多待解決的問題，例如: 龐大的硬體資源使用量、過大的預測錯誤率、過低的執行頻率、以及外界光源雜訊的干擾等等。因此，如何設計一套視差預測演算法，在有限的硬體資源應用下，亦能滿足低錯誤率、高執行頻率、以及抗光噪等要求是本文研究之重點。
在本篇論文當中提出了一種基於引導影像濾波器的視差預測演算法，分別針對 (1)降低硬體資源與提高執行頻率、(2)降低預測錯誤率、(3)降低光噪影響等三大效能指標進行研究。在降低硬體資源與提高執行頻率方面，本設計基於圖像間相互較驗概念，設計左右決策樹檢查法，使左圖與右圖在進行視差計算時，分別只需要計算偶數與奇數視差值，從而將硬體資源減半。在降低預測錯誤率方面，本設計導入連續平面分割法，針對相同平面內的不一致點進行校正，以提昇整體準確率。在降低光噪影響方面，本設計導入漢明距離法概念，在受光線亮度影響的狀態下，依舊能搜尋左右圖像間的相似點，產生正確視差值。
所提出的演算法經由Xilinx Vertex-7與Kintex-7的現場可編程門陣列實現電路，並分別使用Middlebury -version 2與 -version 3數據集對其性能進行了測試，version 2資料集主要針對無光噪影像進行視差準確度測試，version 3資料集則加入了光噪因素，同時在圖像上的複雜度也相對應增加。由實驗結果得知，我們的演算法可達133 fps之運行速度，在version 2 的測試資料集當中平均錯誤率僅為6.36%，相比於基於GIF架構的視差預測演算法，可達到最高的執行頻率與低的測試錯誤率。在version 3的測試資料集當中，平均錯誤率為19.65%，與同樣可應用於光噪環境之下的演算法相比較，亦可達到最低的測試錯誤率，此外，本設計亦是第一個可應用於光噪環境當中的基於GIF架構視差演算法。最終我們將實際環境中所拍攝而得的影像藉由本設計進行視差預測，已可達到優異的預測結果，由此證明本設計具有相當之實用性。

With the advancement of science and technology, human needs for image applications have gradually transformed from traditional two-dimensional images to three-dimensional images. In other words, one more demand for "depth information" is required in the image data. The depth information can be widely used in various forward-looking systems, such as autonomous driving technology, augmented reality (AR), virtual reality (VR), robot navigation judgment. Therefore, designing an effective depth prediction method to generate high-precision depth information is an important and urgent issue.
Among various depth prediction methods, the simplest and most effective way is the dual-lens depth method. Via this method, the left-image and right-image are first taken by two parallel lenses. The disparity estimation algorithm is designed to estimate the disparity values between the left- and right- image. The final depth result is then generated with the known information such as the focal length and the distance between two lenses. Thus, the disparity estimation algorithm can be regarded as the vital step in the dual-lens depth method, and its performance significantly impacts the final depth information results. Many studies have also implemented the disparity estimation algorithm into the digital circuit to improve performance in recent years. However, there are still several problems in practical applications that need to be solved, such as colossal hardware resource usage, excessive prediction error rate, low execution frequency, and interference from external light noise. Therefore, a practical disparity estimation algorithm must be designed to satisfy the requirements for low error rate, high operation frequency, and resistance to light noise under the application of limited hardware resources.
In this dissertation, a guided image filter (GIF) -based disparity estimation algorithm is proposed, which aims at (1) reducing hardware resources and increasing the operation frequency, (2) reducing the prediction error rate, and (3) reducing the impact of light noise. Regarding the first target, based on the concept of mutual confirmation, we adopted the left-right decision tree check method in our algorithm so that the left and right images only check the even and odd disparity values, respectively, thereby halving resource costs. In terms of the second target, our algorithm introduces the continuous plane segmentation method to correct the inconsistent points during the same plane to enhance the overall accuracy. Finally, to achieve the last target, the Hamming distance method is adopted in our algorithm. It can still search for similar points between the left and right images under light noise and generate the correct disparity values.
The proposed algorithm was implemented with the Xilinx Vertex-7 and Kintex-7 field-programmable gate array (FPGA), and its performance was tested using the Middlebury -version 2 and -version 3 data sets. The version 3 data set is used primarily for testing the accuracy of disparity values without light noise. The version 3 data set adds the light noise factor, and the complexity of the images in this data set also increases. According to the experiment results, the proposed algorithm provides an operation speed of 133 fps with an error rate of only 6.36% in the version 2 data set. Our algorithm can achieve the highest operation frequency and lowest testing error rate compared with other GIF-based disparity estimation algorithms. When testing by the version 3 data set, the average error rate is 19.65%, which can also perform the lowest test error rate compared with algorithms that can also be applied to light-noise environments, which is caused by lighting noise. Besides, this design is also the first GIF-based algorithm, which can be applied in the light-noise environment. In the end, we use the proposed algorithm to estimate the disparity values of the images taken in the actual environment and get excellent estimation results, which proves that the proposed algorithm is quite suitable to be applied in practical cases.

[1] M. Gong and Y. Yang, "Real-Time Stereo Matching Using Orthogonal Reliabil-ity-Based Dynamic Programming," in IEEE Transactions on Image Processing, vol. 16, no. 3, pp. 879-884, Mar. 2007.
[2] A. Klaus, M. Sormann and K. Karner, "Segment-Based Stereo Matching Using Belief Propagation and a Self-Adapting Dissimilarity Measure," 18th International Conference on Pattern Recognition (ICPR), Hong Kong, 2006.
[3] Vladimir Kolmogorov, Pascal Monasse, Pauline Tan. Kolmogorov and Zabih, "Graph Cuts Stereo Matching Algorithm," Image Processing On Line (IPOL), vol. 4, pp.220-251, 2014.
[4] B. Tippetts, D. J. Lee, K. Lillywhite, and J. Archibald, "Review of stereo vision al-gorithms and their suitability for resource-limited systems," Journal of Real-Time Image Process., vol. 11, no. 1, pp. 5–25, Jan. 2016.
[5] R. A. Hamzah and H. Ibrahim, "Literature survey on stereo vision disparity map algo-rithms," Journal of Sensors, Art. no. 8742920, 2016.
[6] R. Kalarot, J. Morris, and G. Gimel'farb, "Performance analysis of multi-resolution symmetric dynamic programming stereo on GPU," in Proc. 25th Int. Conf. Image Vis. Comput. New Zealand (IVCNZ), Nov. 2010.
[7] L. Zhang, K. Zhang, T. S. Chang, G. Lafruit, G. K. Kuzmanov, andD. Verkest, "Re-al-time high-definition stereo matching on FPGA," in Proc. 19th ACM/SIGDA Int. Symp. Field Program. Gate Arrays, 2011.
[8] S. Jin et al., "FPGA design and implementation of a real-time stereo vision system," IEEE Transactions on Circuits and Systems for Video Technology, vol. 20, no. 1, pp. 15–26, Jan. 2010.
[9] K.-J. Yoon and I.-S. Kweon, "Adaptive Support-Weight Approach for Correspond-ence Search," IEEE Transactions. Pattern Matching and Machine Intelligence, vol. 28, no. 4, pp. 650-656, Apr. 2006.
[10] F. Tombari, S. Mattoccia, and L. D. Stefano, "Segmentation based adaptive support for accurate stereo correspondence," in Proc. Pacific-Rim Symp. Image Video Tech-nol., Jun. 2006.
[11] N. Chang, T. Tsai, B. Hsu, Y. Chen, and T. Chang, "Algorithm and Architecture of Disparity Estimation With Mini-Census Adaptive Support Weight," IEEE Transactions on Circuits and Systems for Video Technology, vol. 20, no. 6, pp. 792-805, 2010.
[12] Jingting Ding, Jilin Liu, Wenhui Zhou, Haibin Yu, Yanchang Wang, and Xiaojin Gong, "Real-time stereo vision system using adaptive weight cost aggregation approach," EURASIP Journal on Image and Video Processing, 2011.
[13] W. Wang, J. Yan, N. Xu, Y. Wang, and F.-H. Hsu, "Real-time high-quality stereo vi-sion system in FPGA, " IEEE Transactions on Circuits and Systems for Video Tech-nology, vol. 25, no. 10, pp. 1696–1708, Oct. 2015
[14] M. Jin and T. Maruyama, "Fast and accurate stereo vision system on FPGA," ACM Transactions on Reconfigurable Technology and Systems, vol. 7, Art. no. 3, 2014.
[15] Y. Shan et al., "Hardware acceleration for an accurate stereo vision system using mini-census adaptive support region," ACM Transactions. Embedded Computer Sys-tems, vol. 13, Art. no. 132, 2014.
[16] Kaiming He, Jian Sun, and Xiaoou Tang, "Guided Image Filtering," IEEE Transac-tions. Pattern Matching and Machine Intelligence, vol. 35, no. 6, pp. 1397-1409, 2013.
[17] A. Hosni, M. Bleyer, C. Rhemann, M. Gelautz, and C. Rother, "Real-time local stereo matching using guided image filtering," inProc. IEEE Int. Conf. Multimedia Expo, 2011.
[18] C. Ttofis and T. Theocharides, "High-quality real-time hardware stereo matching based on guided image filtering," in Proc. Des.,Autom. Test Eur. Conf. Exhib., Mar. 2014
[19] C. Ttofis, C. Kyrkou, and T. Theocharides, "A low-cost real-timeembedded stereo vision system for accurate disparity estimation basedon guided image filtering," IEEE Transactions on Computers, vol. 65, no. 9, pp. 2678–2693, Sep. 2016.
[20] C.K. Vala, K. Immadisetty, A. Acharyya, C. Leech, V. Balagopal, G.V. Merrett, and B.M. Al-Hashimi, "High-speed low-complexity guided image filtering-based disparity estimation," IEEE Transactions on Circuits and Systems I: Regular Papers, vol. 65, no. 2, pp. 606–617, 2018.
[21] J. Subramaniam, J. K. Raju, and D. Ebenezer, "Fast median-finding word comparator array," Electronic Letter, vol. 53, no. 21, pp. 1402–1404, Dec. 2017
[22] L. F. S. Cambuim, J. P. F. Barbosa and E. N. S. Barros, "Hardware module for low-resource and real-time stereo vision engine using semi-global matching ap-proach," in Symposium on Integrated Circuits and Systems Design (SBCCI), Forta-leza, 2017.
[23] M. Dehnavi, M. Eshghi, "Cost and power efficient FPGA based stereo vision system using directional graph transform, ", Journal of Visual Communication and Image Representation, vol. 56, pp. 106-115, Oct. 2018.
[24] 戴子祐, " Hardware Implementation of low-cost disparity refinement algorithm", 國立成功大學資訊工程研究所,碩士論文, 民國109年
[25] 方宏育, " Hardware Implementation of High-performance low-cost disparity predic-tion algorithm based on regional similarity ", 國立成功大學資訊工程研究所,碩士論文, 民國108年
[26] R.-D. Chen, P.-Y. Chen, and C.-H. Yeh, "Design of an area-efficient one-dimensional median filter," IEEE Transactions on Circuits and Systems II, Express Briefs, vol. 60, no. 10, pp. 662–666, Oct. 2013.
[27] E. Nikahd, P. Behnam and R. Sameni, "High-Speed Hardware Implementation of Fixed and Runtime Variable Window Length 1-D Median Filters," IEEE Transactions on Circuits and Systems II: Express Briefs, vol. 63, no. 5, pp. 478–482, May 2016.
[28] Smith, John L. "Implementing median filters in xc4000e fpgas." Xilinx Xcell 23 (1996): 16.
[29] Vega-Rodríguez, Miguel A., Juan M. Sánchez-Pérez, and Juan A. Gómez-Pulido. "An FPGA-based implementation for median filter meeting the real-time requirements of automated visual inspection systems," Proc. 10th Mediterranean Conf. Control and Automation, 2002.
[30] J. Subramaniam, J. K. Raju, and D. Ebenezer. "Parallel and Pipelined 2-D Median Filter Architecture," IEEE Embedded Systems Letters, vol. 10, no. 3, pp. 69–72, Nov. 2017.
[31] I. C. Lin, W. T. Chen, Y. C. Chou, P. Y. Chen, " A Novel Comparison-Free 1-D Me-dian Filter," IEEE Transactions on Circuits and Systems II: Express Briefs, vol. 67, no. 7, pp. 1329-1333, Jul., 2020.
[32] J. O. Cadenas, G. M. Megson and R. S. Sherratt, "Median Filter Architecture by Ac-cumulative Parallel Counters," IEEE Transactions on Circuits and Systems II: Express Briefs, vol. 62, no. 7, pp. 661–665, Jul. 2015.
[33] S. H. Lin, P. Y. Chen, and C. K. Hsu, "Modular Design of High-Efficiency Hardware Median Filter Architecture," IEEE Transactions on Circuits and Systems I: Regular Papers, vol. 65, no. 6, pp. 1929–1940, June 2018.
[34] 蕭佑哲, " Design and implementation of a low latency dual path sorter ", 國立成功大學資訊工程研究所,碩士論文, 民國108年
[35] W. T. Chen, P. Y. Chen, Y. C. Hsiao, and S. H. Lin, "A Low-Cost Design of 2D Me-dian Filter," in IEEE Access, vol. 7, pp. 150623-150629, Oct. 17, 2019.