Development of Mobile Phone Box Detection Modul Using Convolutional Neural Network for Identification of Stock Smartphone

Determining the percentage of product distribution of a company compared to the products of its competitors is an important thing to know by the company. Through this research, the company can conduct market research on the products it sells. With the development of the existing artificial intelligence system, it is possible to determine the distribution of these products can using Deep Learning. On this occasion will be tried using Convolutional Neural Network-based Object Detection to detect cell phone boxes that scattered on a cellphone counter. Through the dropping of the mobile phone, the box will also know the number of each object or cell phone box that is detected, both from its products and competitors' products, so that the company's product distribution compared with other competitors. The model architecture used in this study is EfficientDet with EffciientNet-B0 Extractor, EffciientNet-B1, EffciientNet-B2, and EffciientNet-B3 features. The evaluation results are EfficientDet-D0 with mAP of 66%, EfficientDetD1 with mAP of 72.5%, EffcientDet-D2 with mAP of 73.8%, EffcientDet-D3 with mAP of 72.7%. Judging from the results of the evaluation. The model and architecture chosen for use in this system are EffcientDet-D0, because the accuracy provided is not too far from other architectures, besides EfficientDet-D0 also has the fastest inference time compared to the others, i.e., 7 ms on GPU RTX 2080 Ti, suitable for use in production with many users.