| Paper Abstract and Keywords |
| Presentation |
2024-06-07 13:15
Lightweight Object Detection Model for a Binary Feature Extractable CMOS Image Sensor
Keiichiro Kuroda, Yudai Morikaku, Yu Osuka (Ritumeikan Univ), Ryuichi Ujiie, Daisuke Morikawa, Hideki Shima (Nisshinbo Micro Devices), Okura Syunsuke, Kota Yoshida (Ritumeikan Univ) |
| Abstract |
(in Japanese) |
(See Japanese page) |
| (in English) |
For the coming Society 5.0, we propose an object detection system using a CMOS image sensor capable of extracting binary feature data in order to reduce power consumption of recognition systems. First, a lightweight deep neural network (DNN) for feature data is verified based on YOLOv7. Despite a decrease in object recognition accuracy of large objects (APL50) by only 4.6% compared to the YOLOv7 model for color images, the DNN parameters, FLOPs, and GPU power consumption are reduced by 66.1%, 76.7%, and 32.5%, respectively. Secondly, a CMOS image sensor capable of extracting binary feature data is proposed. Simulation results demonstrate that, while the APL50 is reduced by 23.1% compared to RGB color images, the image sensor output data is reduced by 99.4% with run-length encoding. |
| Keyword |
(in Japanese) |
(See Japanese page) |
| (in English) |
CMOS image sensor / feature extraction / object detection / data reduction / Run Length Encoding / / / |
| Reference Info. |
ITE Tech. Rep., vol. 48, no. 17, IST2024-27, pp. 23-28, June 2024. |
| Paper # |
IST2024-27 |
| Date of Issue |
2024-05-30 (IST, ME) |
| ISSN |
Online edition: ISSN 2424-1970 |
| Download PDF |
|
|