Google Vision的問題，透過圖書和論文來找解法和答案更準確安心。 我們找到下列各種有用的問答集和懶人包

Math and Architectures of Deep Learning

為了解決Google Vision的問題，作者Chaudhury, Krishnendu 這樣論述：

Krishnendu Chaudhury is a deep learning and computer vision expert with decade-long stints at both Google and Adobe Systems. He is presently CTO and co-founder of Drishti Technologies. He has a PhD in computer science from the University of Kentucky at Lexington.

Google Vision進入發燒排行的影片

應用於物件偵測與關鍵字辨識之強健記憶體內運算設計

為了解決Google Vision的問題，作者江宇翔 這樣論述：

近年來，由於不同的應用都能夠藉由和深度學習的結合而達到更好的結果，像是物件偵測、自然語言處理以及圖像辨識，深度學習在終端設備上的發展越來越廣泛。為了應付深度學習模型的龐大資料搬移量，記憶體內運算的技術也在近年來蓬勃發展，不同於傳統的范紐曼架構，記憶體內運算使用類比域的計算使儲存設備也同樣具備運算的能力。儘管記憶體內運算具有降低資料搬移量的優點，比起純數位的設計，在類比域進行計算容易受到非理想效應的影響，包括元件本身或是周邊電路的誤差，這會造成模型災難性的失敗。此篇論文在兩種不同的應用領域針對記憶體內運算進行強健的模型設計及硬體實現。在電阻式記憶體內運算的物件偵測應用當中，我們將重點放在改善模

型對於非理想效應的容忍度。首先，為了降低元件誤差的影響，我們將原本的二值化權重網路改變為三值化權重網路以提高電阻式記憶體中高阻態元件的數量，同時能夠直接使用正權重及負權重位元線上的電流值進行比較而不使用參考位元線作為基準。其次，為了避免使用高精度的正規化偏差值以及所導致的大量低阻態元件佈署，我們選擇將網路中的批次正規化層移除。最後，我們將運算從分次的電流累加運算改為一次性的運算，這能夠將電路中非線性的影響降到最低同時避免使用類比域的累加器。相較於之前的模型會受到這些非理想效應的嚴重影響導致模型無法運作，我們在考慮完整的元件特性誤差，周邊電路誤差以及硬體限制之下，於IVS 3cls中做測試，能夠

將平均精確度下降控制在7.06\%，在重新訓練模型後能更進一步將平均精確度下降的值降低到3.85\%。在靜態隨機存取記憶體內運算的關鍵字辨識應用當中，雖然非理想效應的影響相對較小，但是仍然需要針對周邊電路的誤差進行偏壓佈署補償，在經過補償及微調訓練後，在Google Speech Command Dataset上能夠將準確率下降控制在1.07\%。另外，由於語音訊號會因為不同使用者的資料而有大量的差異，我們提出了在終端設備上進行模型的個人化訓練以提高模型在小部分使用者的準確率，在終端設備的模型訓練需要考量到硬體精度的問題，我們針對這些問題進行誤差縮放和小梯度累積以達到和理想的模型訓練相當的結果

。在後佈局模擬的結果中，這個設計在推論方面相較於現有的成果能夠有更高的能源效率，達到68TOPS/W，同時也因為模型個人化的功能而有更廣泛的應用。

Korean Food Systems: Secrets of the K-Diet for Healthy Aging

為了解決Google Vision的問題，作者 這樣論述：

Dr. Kalidas Shetty is currently the Associate Vice President for International Partnerships & Collaborations and Founding Director of Global Institute of Food Security and International Agriculture-GIFSIA & Professor of Plant Sciences at North Dakota State University, Fargo, ND, USA. Dr. Shetty’s re

search interests focus on critical role of cellular and metabolic basis of oxygen biology for advancing new innovations in Life Sciences and especially Agricultural and Food Innovations that advance global food security and health in a sustainable environment. His specific research interests focus o

n scientific, educational and policy strategies to advance climate resilient health-targeted food security solutions including malnutrition and hunger challenges. In particular, he has developed an innovative climate resilient crops for health research platform to counter diet-linked non-communicabl

e chronic diseases (NCD). He has published over 225 manuscripts in peer-reviewed journals and over 50 as invited reviews and in conference proceedings with H-Index of 75 on Google Scholar. In 2004, he was selected by US State Department as the inaugural Jefferson Science Fellow to advice on scientif

ic issues as it relates to International Diplomacy and International Development. Dr. Shetty has widely traveled and has been invited to present lectures and seminars in the areas of Food Biology, Climate Resilient Healthy Food Systems for Food Security & Health and Food Safety in over 50 countries

in Asia, Europe, Africa, and the Americas. He also has deep commitment to global education capacity building.His current passion is to advance research, education capacity building and policy on sustainable and ecological basis for climate resilient healthy food systems and food diversity to drive g

lobal food security. This vision is based on crops and food diversity, indigenous wisdom, traditional fermentations, and new technology innovations in ethnic and indigenous food systems that incorporates understanding of comparative cellular biochemistry of plant and animal systems and their interac

tions with microbial systems in diverse ecologies and cultures of earth. This system based integrated model and research platform based on cellular basis of oxygen biology of food plants and plant-microbial interactions is the basis for new and innovative sustainable agriculture and food solutions t

o advance climate resilient and health-targeted food security. Dong-Hwa Shin, PhD is an emeritus professor of Chonbuk National University in Jeonju, Korea. He has his Ph. D in Food Science & Technology at Dongguk University, Seoul, Korea, He has 18 years’ experience in food research field at Food Re

search Institute in AFDC which is first Korean government supported food research organization. At this institute he carried out many industries-oriented research and applied research. After working at the institute, he transferred to Chonbuk National University in Dept. of Food Science & Technology

and served for 22 years.During this time, he published 350 research papers, 15 patents and 23 books. He managed a Regional Research Center (RRC) for supporting food industry in the region which was fully supported by Korean government for 10 years. His major field is food fermentation, especially s

oybean and vegetable-based fermentation. Based on his research he transferred 30 relevant techniques to the related industries to be commercialized. He had served as advisor of different Governmental organizations and private food manufacturing companies including CJ and Nongshim, which are the bigg

est companies in Korea. He was a UNDP/FAO consultant for food industry development at Southeast Asian countries (1983). He is a member of Korean Academy of Science & Technology (KAST), the president of Korean Society of Food Science Technology (2002), President of Korean Society of Food Hygiene and

Safety, President of Korean Association of Food Sanitation (2004-2016). and President of Food Safety Association. At present he is the president of Korean Council for Sunchang Soybean Fermentation, the president of Korea Food Industry Promotion Forum, and running Shindonghwa Food Research Institute

as Director.

基於FPGA單晶片及像素趨勢車道線檢測法實現車道線感測系統之研究

為了解決Google Vision的問題，作者廖國欽 這樣論述：

車輛自動駕駛系統目前主要是由自動跟車 (Adaptive Cruise Control, ACC) 以及車道偏離警示 (Lane Departure Warning System, LDWS) 兩大系統所組成。然而，自動跟車系統在實現過程中，由於必須藉由前方車輛實現車輛跟隨功能，因此若無前方車輛時則無法實現此功能。反觀車道偏離警示系統是依據車道線軌跡來幫助車輛保持於車道內，因此具備較高實用性。在此，本研究特別針對車道感測進行研究。由於傳統的車道線感測必須仰賴高效率的電腦才能有效地完成運算，為了克服傳統車道線辨識的缺點，本研究專注於如何將車道線辨識演算法簡化，並實現在單晶片上，達到低功耗之目的

。本研究以單一數位相機及單一現場可程式邏輯閘陣列 (Field Programmable Gate Array, FPGA) 實線以精簡之硬體電路達到即時於白天及黃昏情況下進行車道線辨識。透過像素趨勢車道檢測法 (Pixel Trend Lane Detection, PTLD) 擷取特徵，並將所得之車道位置利用線性回歸 (Linear Regression, LR) 決定車道線的軌跡，再透過左右車道回歸線取得車道的中心線，藉此引導車輛穩定行駛於車道中。另外，本研究還搭配語音辨識擴充模組 (DFR0177 Voice Recognition) 來辨識由Google Map路線規劃所傳出的語音指

令。根據辨識的結果，輸出行車指令給FPGA，以此決定車輛轉彎或直線行車路線模式。根據本研究之實驗結果，在使用每秒90張畫面播放速度以及640×480影像解析度情況下，只需11 ms即可擷取車道線特徵。而由左右車道線線性回歸決定出的中心線與實際影像中的中心線，誤差僅在5個像素以內。故本研究不管在運算速度以及準確度上均符合實際運用需求，未來可以有效幫助車輛穩定行駛於車道，達成自動駕駛之目的。