Optimal Object Classification Model for Embedded Systems based on Pre-trained Models

• Published in: 2021 25th International Computer Science and Engineering Conference (ICSEC) pp. 307 - 312
• Main Authors: Thokrairak, Sorawit, Thibuy, Kittiya, Fongsamut, Chalermpan, Jitngernmadan, Prajaks
• Format: Conference Proceeding
• Language: English
• Published: IEEE 18-11-2021
• Subjects: Classification; COCO; Computational modeling; Computer science; Embedded System; Embedded systems; Graphics processing units; Machine learning; Random access memory; Sorting; SSD-MobileNet
• DOI: 10.1109/ICSEC53205.2021.9684656

Machine Learning technology grows in the field of automatic waste sorting machines equipped with an intelligent unit. This intelligent unit runs on an embedded system that mostly has lower computation power (both CPU and GPU) and lower RAM. However, to archive a higher accuracy rate on classification one has to use a sophisticated classification AI model, which needs less computational power. We considered our experiment using 3 pre-trained models based on COCO dataset, namely the ssd_mobilenet_v2_coco, the ssd_inception_v2_coco, and the ssd_resnet_50_fpn_coco due to their good quality. The results show that although the AI model based on the ssd_resnet_50_fpn_coco has the highest accuracy (99.75%), it consumes the most computational power. In contrast, the one based on the ssd_mobilenet_v2_coco has acceptable accuracy (98.83%) and it consumes the lowest computational power. We decided that the most suitable AI model for embedded systems is the one that is trained with the pre-trained ssd_mobilenet_v2_coco model.