TY - GEN
T1 - Brand Recognition with Partial Visible Image in the Bottle Random Picking Task based on Inception V3
AU - Zhu, Chen
AU - Matsumaru, Takafumi
N1 - Funding Information:
*This research is supported by Japan Society for The Promotion of Science (KAKENHI-PROJECT-17K06277), to which we would like to express our sincere gratitude.
Publisher Copyright:
© 2019 IEEE.
PY - 2019/10
Y1 - 2019/10
N2 - In the brand-wise random-ordered drinking PET bottles picking task, the overlapping and viewing angle problem makes a low accuracy of the brand recognition. In this paper, we set the problem to increase the brand recognition accuracy and try to find out how the overlapping rate infects on the recognition accuracy. By using a stepping motor and transparent fixture, the training images were taken automatically from the bottles under 360 degrees to simulate a picture taken from viewing angle. After that, the images are augmented with random cropping and rotating to simulate the overlapping and rotation in a real application. By using the automatically constructed dataset, the Inception V3, which was transferred learning from ImageNet, is trained for brand recognition. By generating a random mask with a specific overlapping rate on the original image, the Inception V3 can give 80% accuracy when 45% of the object in the image is visible or 86% accuracy when the overlapping rate is lower than 30%.
AB - In the brand-wise random-ordered drinking PET bottles picking task, the overlapping and viewing angle problem makes a low accuracy of the brand recognition. In this paper, we set the problem to increase the brand recognition accuracy and try to find out how the overlapping rate infects on the recognition accuracy. By using a stepping motor and transparent fixture, the training images were taken automatically from the bottles under 360 degrees to simulate a picture taken from viewing angle. After that, the images are augmented with random cropping and rotating to simulate the overlapping and rotation in a real application. By using the automatically constructed dataset, the Inception V3, which was transferred learning from ImageNet, is trained for brand recognition. By generating a random mask with a specific overlapping rate on the original image, the Inception V3 can give 80% accuracy when 45% of the object in the image is visible or 86% accuracy when the overlapping rate is lower than 30%.
UR - http://www.scopus.com/inward/record.url?scp=85078859351&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85078859351&partnerID=8YFLogxK
U2 - 10.1109/RO-MAN46459.2019.8956374
DO - 10.1109/RO-MAN46459.2019.8956374
M3 - Conference contribution
AN - SCOPUS:85078859351
T3 - 2019 28th IEEE International Conference on Robot and Human Interactive Communication, RO-MAN 2019
BT - 2019 28th IEEE International Conference on Robot and Human Interactive Communication, RO-MAN 2019
PB - Institute of Electrical and Electronics Engineers Inc.
T2 - 28th IEEE International Conference on Robot and Human Interactive Communication, RO-MAN 2019
Y2 - 14 October 2019 through 18 October 2019
ER -