Experience

由田新技股份有限公司

• Appliances, Electrical, and Electronics Manufacturing
• 1 - 100 Employee

• Algorithm Engineer

  • Oct 2019 - Present

  負責新進軟體工程師訓練、實作CUDA演算法加速dll、對新和舊演算法做公式優化和為各部門客製化實現數值或統計的演算法功能。 教授新人的訓練課程含括檢測機台的控制、基礎影像處理、C#和多執行緒等實機操作。完成GPU加速的既有演算法有: skeleton(thinning ZS and GH Methods)、1-8 bit image transformation、geometric transformation、Matching(PCC)、morphological operation、convolution和FFT。另外也針對線路和圖像檢測做GPU/CPU上的速度優化，利用CPU擷取必要資訊時，同時在GPU計算上一步的結果來節省時間。除了Geometric校正，也實作將非線性和線性fitting完的參數使用GPU來做圖像校正。值得一提的有將修正過後的skeleton演算法在GPU上執行能比在CPU上存在已知bug的原版skeleton版本還快上20~240倍的速度。 開發和使用的平台主要為Windows，並使用Ubuntu來做多CPU/GPU的驗證。使用到的語言、套件或工具有: un/managed C++、C#、CUDA、NPP、cuFFT、SSE/AVX和OpenCV。 I am responsible for trainning new software engineers, implementing algorithm acceleration dlls on CUDA, optimizing calculate formulas in algorithms with both novel and traditional algorithms, and customizing numerical or statistics algorithms for different product groups. The trainning for new engineers include programming control systems on inspection machines, basic image processing, C#, multi-thread, and operating demos on real inspaction machines. For GPU acceleration parts, I have accomplished the following well known algorithms: skeleton(thinning ZS and GH Methods), 1-8 bit image transformation, geometric transformation, Matching(PCC), morphological operation, convolution, and FFT. In existing image and circuit inspction algorithms, I successfully saved up to 50% time cost by hiding calculating results on last step with gpu background when cpu getting image info in the current step. Besides geometric transformation, I also implemented non/linear fitting correction methods for distortions images with gpu and cpu. Worth mentioning, the speedup of improved skeleton algorithms on GPU is around 20 to 240 times faster than original methods which exist well known bug on CPU. The main develope and deploy enviroment is Windows. The functions will also be tested on Ubuntu for verifications of multi CPU/GPU parallel computing. The tools used in current position are: un/managed C++, C#, CUDA, NPP, cuFFT, SSE/AVX, and OpenCV. Show less



WiBase

• 台灣 台北市

• Software Engineer

  • Mar 2018 - Oct 2019

  主要從事開發NVidia的Tegra系列產品的軟體應用。此外，也和數十家純軟體公司做軟體技術支援和交流。獨立製作各種基於Nvidia的機器學習、IVA、電腦視覺和AI神經網路的完整展示，對產品開發展示和方案可行性提供方向。時常需要和世界各國的軟體公司或SI公司直接con-call討論交流。另外需要協助NVidia相關的交流或研討會(如GTC、Computex等)，代表公司向軟體相關產業洽談可能的合作。 使用平台為Ubuntu(PC)、Ubuntu(ARM)、Raspbian、Arduino、Windows10和Windows IOT。主要開發語言為C/C++、CUDA、C#。使用的軟體套件和SDK框架等包含Gstreamer、Deepstream、OpenCV、NV multimedia、Caffe、TensorFlow。 做過的幾個主要專案: 1. 自製並訓練小型模型以利RAW攝影機以120fps做即時撲克牌辨識。使用C++、Caffe、Gstreamer和OpenCV實作於TX2上。其中關鍵是利用硬體加速(非CUDA)將影像轉換格式，並將其傳入gstreamer plugin，以便快速做模型辨識。 2. 和電機機構同事一起架設四個FOV為90度以上的攝影機於車頂，將左右前和左右後攝像頭拼接後，並將TX2搬至車上做即時辨識。此系統讀取車用OBD配上TX2上之GPS和網路開源之地圖做即時HUD顯示，也有裝上光達作測距。另外再加上物件辨識(車輛、號誌、行人)和車道線辨識(使用CUDA)，以模擬後裝之近ADAS系統。實作使用C++、Deepstream、Gstreamer、Caffe、OpenCV和硬體加速。正確的顯示即時解析完的GPS和OBD資料蠻有挑戰性的。 3. 拆解市售遙控車，並於其裝上Jetson nano、Arduino和USB攝像頭。將畫面由nano經WIFI傳出RTSP串流給遠端電腦程式(C#)或是手機，接著由遠端下達移動指令給nano，再經由RJ45傳入Arduino做pin腳電訊號控制遙控車本身的機電系統。三個開發平台上除了使用到C++、Deepstream、Gstreamer，也實作gstreamer plugin做RTSP串流和socket封包溝通做pin腳電位控制。 4. 在和眾多軟體上的合作案中，最值得提的是和NEC臉部辨識的軟體socket接口測試案。NEC的臉部辨識系統是運作在Windows上的臉部特徵辨識系統，其需要edge端發送人臉至其辨識系統，辨識後再將結果傳回edge端，做接續的顯示或是門禁控制。edge端在TX2上實作，基於Deepstream上修改了plugin將其結果篩出並用socket傳出Json格式的base64圖檔。內部TX2壓力測試的情況下，可以同時6隻FHD@30fps攝影機的影像做辨識並丟出每秒達1800張人臉(180pixel*240pixel)，並同時顯示。 Mainly focus on developing applications on Tegra platforms from NVidia. I also cooperate projects with, and provide technical support to dozens of software companies. In the fields of ML, IVA, CV, and deep learning, I promote and demonstrate the possibility of new solutions on NVidia platform alone. As a part of my job, I also need to act like a consultant in AI solutions. In most of committees related to NVidia, they are also my responsibility to take in charge of software part, such as GTC or CES. Develop OS(platform) are Ubuntu(x86/ARM), Windows(10/IoT), Raspbian, and Arduino with C/C++, CUDA, C#. The frameworks and SDKs often used in projects are Gstreamer, Deepstream, OpenCV, NV multimedia, Caffe, and Tensorflow. Show less



捷螺系統股份有限公司

• 新竹市

• Algorithm Engineer

  • Oct 2017 - Mar 2018

  主要使用機器學習方式開發藍芽和WIFI的室內定位演算法。另外也需要協同前端做後端的分散式運算。使用到的平台軟體為在VisualStudio之C/C++、C#和Node.js。 Develop indoor-positioning algorithm with machine learning based on Bluetooth and WIFI signals. It is also necessary to cooperate with front-end and implement small scale of distributed computing in back-end. The programing languages are C/C++, C#, and Node.Js. 主要使用機器學習方式開發藍芽和WIFI的室內定位演算法。另外也需要協同前端做後端的分散式運算。使用到的平台軟體為在VisualStudio之C/C++、C#和Node.js。 Develop indoor-positioning algorithm with machine learning based on Bluetooth and WIFI signals. It is also necessary to cooperate with front-end and implement small scale of distributed computing in back-end. The programing languages are C/C++, C#, and Node.Js.