Research and develop algorithms in my company. Cooperate with domestic and foreign manufacturers to use AI technology to solve problems that traditional algorithms cannot overcome. Completed or ongoing projects include defect detection, text recognition (OCR), airplane detection, random bin picking, bottle inspection, point cloud image cutting, and robotic arm grasping by reinforcement learning.
Cellphone: +886 0952152828
Gmail: [email protected]
Personal Web: https://www.ivan-lee.me/
Blog: https://medium.com/change-the-world-with-technology
Recruit AI researchers and RD. Planning and project management.
1. Plan project development and project process
2. Organize the paper-sharing club
3. Plan coding style python with team members
4. Lead newcomers to familiarize the project and development
Research and developing visual recognition and control, using AI to solve multi-dimensional feature problems that cannot be handled by traditional methods, and making the algorithm to achieve generalization and desired speed.
1. Develop vision algorithms
2. Research control algorithms of robotic-arm
3. Deploy algorithms to embedded machines (Nvidia AGX, Nvidia NX)
4. Track the latest cutting-edge journals and technologies
Tensorflow、AlexNet、Cross entropy method
Traditional algorithms can only specify a single or few items after modeling. Using an AI algorithm, it can grasp various daily necessities or stationery (universal). As long as the width of the object gripper and a specific distance are met, 95% of daily life and household items can be grasped. In cooperation with the robotic arm team, the company's project has been able to grab a variety of daily necessities, including fried chicken legs and bundled wires (flexible objects).
Tensorflow、Segmentation、Unet
Cooperate with well-known factories and manufacturers to detect the defects of various items. Traditional algorithms need to design multiple layers of logic and processing for complex images, but general-purpose neural networks can overcome this problem. Design a general-purpose model, so that the model (Segmentation) can be generalized and learned effectively. The objects include 17 kinds of items, such as tires, keyboards, PCB boards, metal welding objects, passive components, etc. Compared with traditional algorithms, it can effectively reduce development by 80%.
Tensorflow、Cross entropy method、Pybullet、Q-learning
The traditional algorithm relies on line scanning and surface scanning. After obtaining the point cloud, it performs a clamping calculation, path planning, and collision detection. The neural-like control method can directly use the color camera, save the expensive point cloud camera, and save all the aforementioned calculation processes, and dynamically execute the grasping strategy. At present, research in the virtual environment has achieved results. After transferring the model to the real world, it is expected to save 25% of the hardware cost of the gripping project and speed up the gripping time by 3 times. demo video:https://youtube.com/shorts/17ROS385zy4
Pytorch、Yolov3、Jetson AGX、Jetson NX
Cooperate with a large domestic institution to detect aircraft on satellite images and actually deploy them on embedded machines after training.
Pytorch、OCR、crnn、scikit-image
Cooperate with well-known domestic retailers to test the nutrition labels and ingredients on bento boxes. The text recognition software on the market cannot detect special Chinese characters (words for nutrition), and the arrangement is too narrow. Therefore, according to the font used on the label, and the image characteristics of the actual scene, such as deformation and skewness, we provide customized products for customers. And passed the customer stress test: in addition to normal words, it can also issue warnings when there are defects or typos in the text, with an accuracy of 99%.
Tensorflow、Segmentation、Edge detection、Surface rebuild、Open3d
Transparent objects have always been a difficult problem in traditional algorithm detection. No matter the line scan, area scan, or depth camera, there will be refraction and transmission, and complete imaging cannot be achieved. Taking advantage of the neural-like feature of processing multi-dimensional information, a variety of models are used to restore the smooth point cloud on the transparent surface of the bottle, overcoming the problem that traditional algorithms cannot solve.
Tensorflow、Mask RCNN
In the traditional method of point cloud calculation, the calculation time is too time-consuming, and the calculation is slow for the final prediction and grasping of traditional CAD. Combined with Instance segmentation to calculate point cloud, the efficiency is accelerated by 2~4 times.
2022/7 - present
Taipei, Taiwan
Research for reinforcement learning, robotic arm, computer vision, GAN, VR.
2017/2 - 2017/8
Learning Hadoop distributed systems, database planning, Linux operation, agile management.
2012 - 2016
Organization and project management.
Research and develop algorithms in my company. Cooperate with domestic and foreign manufacturers to use AI technology to solve problems that traditional algorithms cannot overcome. Completed or ongoing projects include defect detection, text recognition (OCR), airplane detection, random bin picking, bottle inspection, point cloud image cutting, and robotic arm grasping by reinforcement learning.
Cellphone: +886 0952152828
Gmail: [email protected]
Personal Web: https://www.ivan-lee.me/
Blog: https://medium.com/change-the-world-with-technology
Recruit AI researchers and RD. Planning and project management.
1. Plan project development and project process
2. Organize the paper-sharing club
3. Plan coding style python with team members
4. Lead newcomers to familiarize the project and development
Research and developing visual recognition and control, using AI to solve multi-dimensional feature problems that cannot be handled by traditional methods, and making the algorithm to achieve generalization and desired speed.
1. Develop vision algorithms
2. Research control algorithms of robotic-arm
3. Deploy algorithms to embedded machines (Nvidia AGX, Nvidia NX)
4. Track the latest cutting-edge journals and technologies
Tensorflow、AlexNet、Cross entropy method
Traditional algorithms can only specify a single or few items after modeling. Using an AI algorithm, it can grasp various daily necessities or stationery (universal). As long as the width of the object gripper and a specific distance are met, 95% of daily life and household items can be grasped. In cooperation with the robotic arm team, the company's project has been able to grab a variety of daily necessities, including fried chicken legs and bundled wires (flexible objects).
Tensorflow、Segmentation、Unet
Cooperate with well-known factories and manufacturers to detect the defects of various items. Traditional algorithms need to design multiple layers of logic and processing for complex images, but general-purpose neural networks can overcome this problem. Design a general-purpose model, so that the model (Segmentation) can be generalized and learned effectively. The objects include 17 kinds of items, such as tires, keyboards, PCB boards, metal welding objects, passive components, etc. Compared with traditional algorithms, it can effectively reduce development by 80%.
Tensorflow、Cross entropy method、Pybullet、Q-learning
The traditional algorithm relies on line scanning and surface scanning. After obtaining the point cloud, it performs a clamping calculation, path planning, and collision detection. The neural-like control method can directly use the color camera, save the expensive point cloud camera, and save all the aforementioned calculation processes, and dynamically execute the grasping strategy. At present, research in the virtual environment has achieved results. After transferring the model to the real world, it is expected to save 25% of the hardware cost of the gripping project and speed up the gripping time by 3 times. demo video:https://youtube.com/shorts/17ROS385zy4
Pytorch、Yolov3、Jetson AGX、Jetson NX
Cooperate with a large domestic institution to detect aircraft on satellite images and actually deploy them on embedded machines after training.
Pytorch、OCR、crnn、scikit-image
Cooperate with well-known domestic retailers to test the nutrition labels and ingredients on bento boxes. The text recognition software on the market cannot detect special Chinese characters (words for nutrition), and the arrangement is too narrow. Therefore, according to the font used on the label, and the image characteristics of the actual scene, such as deformation and skewness, we provide customized products for customers. And passed the customer stress test: in addition to normal words, it can also issue warnings when there are defects or typos in the text, with an accuracy of 99%.
Tensorflow、Segmentation、Edge detection、Surface rebuild、Open3d
Transparent objects have always been a difficult problem in traditional algorithm detection. No matter the line scan, area scan, or depth camera, there will be refraction and transmission, and complete imaging cannot be achieved. Taking advantage of the neural-like feature of processing multi-dimensional information, a variety of models are used to restore the smooth point cloud on the transparent surface of the bottle, overcoming the problem that traditional algorithms cannot solve.
Tensorflow、Mask RCNN
In the traditional method of point cloud calculation, the calculation time is too time-consuming, and the calculation is slow for the final prediction and grasping of traditional CAD. Combined with Instance segmentation to calculate point cloud, the efficiency is accelerated by 2~4 times.
2022/7 - present
Taipei, Taiwan
Research for reinforcement learning, robotic arm, computer vision, GAN, VR.
2017/2 - 2017/8
Learning Hadoop distributed systems, database planning, Linux operation, agile management.
2012 - 2016
Organization and project management.