行動裝置與多媒體技術近年來迅速普及，手機取代了傳統底片式相機，成為人們隨手拍照的工具。人們每一年用手機拍攝十二兆張照片，這些照片絕大多數是以JPEG格式壓縮，保存在手機中由快閃記憶體(NandFlash)構成的固態儲存，例如，SD卡及eMMC晶片。它具有體積小、耗電低、速度快、抗撞擊等優點，非常適合用在手機上。隨著製程微縮、垂直堆疊等技術發展，快閃記憶體的儲存密度及容量不斷上升，使得快閃記憶體資訊保存的可靠度降低。經歷長時間儲存之後，存在其中的資訊會因為快閃記憶體的漏電問題產生錯誤。雖然在快閃記憶體中通常會搭配錯誤更正碼修復錯誤，但是錯誤更正碼更正錯誤的能力有一定的上限。因此本論文預想到一個未來潛在的問題：若使用者如果多年沒使用一手機，儲存在手機eMMC晶片及SD卡中的資料會逐年累積retention errors，最終將超過錯誤更正碼能更正的上限，造成使用者的照片毀損。在本篇論文中，我提出當JPEG圖片存入快閃記憶體中的保護及修復策略，在JPEG圖片有毀損的情況產生時，我所提出的機制可以盡量修復圖片。透過將JPEG圖片存入TLC型式的快閃記憶體，我分析10年間資訊錯誤的情況。然後我根據分析所得的位元錯誤率，在JPEG圖片檔案中隨機加入位元錯誤以模擬JPEG圖片毀損時的狀況，再以本論文提出的機制修復。實驗的結果顯示我的機制最高可使PSNR平均提升29 dB，而我的機制所額外記錄的冗餘資訊為13%。

Mobile devices and multimedia technologies have rapidly spread in recent years, and smartphones have replaced traditional film-type cameras and become the most popular tools for people to take pictures.People use smartphones to take 1.2 trillion photos per year. Most of these photos are compressed in the JPEG format and stored in the solid-state storage consisting of NAND flash (flash for short) of smartphones, such as SD cards and eMMC chips. Flash-based storage has the advantages of small form factor, low power consumption, high performance, and shock resistance and is a good fit for smartphones.However, with process shrinking, storing multiple bits per cell, and vertical stacking, the reliability of flash-based storage is decreasing.That is, a significant number of errors appear in the data stored in flash-based storage because of long-term charge leakage of flash storage cells. Although error correction codes (ECCs) are typically used in flash-based storage to handle errors, they all have an error correction limit. Therefore, this paper foresees a potential issue in the future: if users do not use a smartphone for many years, retention errors will accumulate in the data stored in the eMMC chip and SD card of the smartphone year after year and will eventually exceed the limit of the ECC. At that point, the photos stored in the smartphone are likely corrupted. To address this potential issue, in this paper, I propose protecting and repairing strategies for JPEG images in flash-based storage. Once a JPEG image is corrupted, the proposed strategies can help users to restore the image in a best effort manner. I analyze the bit error rate (BER) of TLC (triple-level cell) flash during 10 years of retention time. I inject errors into JEPG files according to the BER and then repair the JPEG files using the proposed mechanisms. The experimental results show that the proposed mechanisms can increase the PSNR of JPEG files by an average of 29 dB, and the storage overhead of the proposed mechanisms is 13%.

誌謝...1
摘要...2
Abstract...3
1 Introduction...8
2 Background and Related Works...12
2.1 快閃記憶體的Retention Errors...12
2.2 JPEG壓縮步驟及流程...13
2.2.1 JPEG圖的標頭檔資訊...14
2.3 圖片質量評估...16
2.4 Related Works...16
3 Design...17
3.1 JPEG圖增加Restart Marker的效益與錯誤蔓延影響...18
3.2 弱Pages的Markers保護...19
3.3 以相似圖像修復存在弱Pages的圖像資訊...21
3.3.1 找尋相似像素區塊的流程...22
3.4 Comment Marker資訊增加對圖片資訊存在強弱Page的影響...24
3.5 修復錯誤的JPEG圖...25
3.5.1 修補弱Pages上解碼錯誤的圖像區塊...26
4 Experiments...28
4.1 實驗環境設置...28
4.2 修復後的JPEG圖片與原圖的PSNR與SSIM結果分析...30
4.3 壓縮率高與壓縮率低的圖片SSIM與PSNR結果分析...32
4.4 單張JPEG圖片內容分析...33
4.5 3個冗餘資訊大小與原圖大小的比較...34
5 Conclusion...38
References...39

[1]C. Cakebread, “People will take 1.2 trillion digital photos this year—thanksto smartphones.” https://www.businessinsider.com/12-trillion-photos-to-be-taken-in-2017-thanks-to-smartphones-chart-2017-8, Aug. 2017.
[2]Y. Cai, E. F. Haratsch, O. Mutlu, and K. Mai, “Error patterns in mlc nandflash memory: Measurement, characterization, and analysis,” inProceedingsof the Conference on Design, Automation and Test in Europe, DATE ’12, (SanJose, CA, USA), pp. 521–526, EDA Consortium, 2012.
[3]Y. Cai, Y. Luo, E. F. Haratsch, K. Mai, and O. Mutlu, “Data retention inmlc nand flash memory: Characterization, optimization, and recovery,” in2015 IEEE 21st International Symposium on High Performance ComputerArchitecture (HPCA), pp. 551–563, Feb 2015.[4]Y. Cai, G. Yalcin, O. Mutlu, E. F. Haratsch, A. Cristal, O. S. Unsal, andK. Mai, “Flash correct-and-refresh: Retention-aware error management forincreased flash memory lifetime,” in2012 IEEE 30th International Conferenceon Computer Design (ICCD), pp. 94–101, Sep. 2012.
[5]A. Sampson, J. Nelson, K. Strauss, and L. Ceze, “Approximate storage insolid-state memories,”ACM Trans. Comput. Syst., vol. 32, pp. 9:1–9:23, Sept.2014.
[6]Q. Guo, K. Strauss, L. Ceze, and H. S. Malvar, “High-density image storageusing approximate memory cells,”SIGARCH Comput. Archit. News, vol. 44,pp. 413–426, Mar. 2016.
[7]E. Yan, K. Zhang, X. Wang, K. Strauss, and L. Ceze, “Customizing progressiveJPEG for efficient image storage,” in9th USENIX Workshop on Hot Topicsin Storage and File Systems (HotStorage 17), (Santa Clara, CA), USENIXAssociation, 2017.
[8]Z. Liu, T. Liu, J. Guo, N. Wu, and W. Wen, “An ecc-free mlc stt-ram basedapproximate memory design for multimedia applications,” in2018 IEEE Com-puter Society Annual Symposium on VLSI (ISVLSI), pp. 142–147, July 2018.