Inpainting Structure of Object by Mask-cycle GAN

• Main Authors: Chan, Chiao-Tung, 詹巧同
• Other Authors: Hu, Jwu-Sheng
• Format: Others
• Language: zh-TW
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/5tm62f

碩士 === 國立交通大學 === 電控工程研究所 === 107 === With the rapid development of technology, the concept and implementation of smart factories have gradually emerged. Furthermore, the application of robotic vision has become more and more extensive, and the demand is also growing. Because of the great progress in the development of AI technology, it replaces traditional mathematical calculations and parameter adjustments, and relies on CNN to achieve better results. The combination of a robotic arm and a camera can help pick up objects in the factory and replace expensive depth sensors with Mask-RCNN RBG data, but stacked objects that are common in factories are often more complex and more difficult to identify than daily photos. Also because of the need for more precise positioning in the factory, general object segmentation is not enough. Another more important reason is that the environmental shielding rate of stacked objects is often very high. Here we have experimented with the inpainting algorithm, and also proposed a method combining mask-RCNN and multiple GANs to solve a single object in a random bin picking environment. The problem of positioning is to use these architectures to restore the pose of the basic object in 3D. This experiment uses the simulation environment Assasim developed by the laboratory and itri team to produce data and verification.