Industrial Automatic Control Systems and Controllers Annotation << Back Rationale of the Choice of the Optimal Neural Network Architecture for the Recognition of Madi History Museum Exhibits from a Mobile Device Camera K.A. Barinov, K.V. Mikolaichuk The paper provides a rationale of the choice of neural network architecture for object detection for further use in the mobile application “Guide to the MADI History Museum”. We consider the principles of constructing architectures of neural networks for object detection and give two modern approaches for constructing architectures, which is «Region-based» and «Single shot» architectures. Taking into account the specifi cs of the task and the necessity of deployment a neural network on a mobile device, we decide to use a detector with «Single shot» architecture. Also an empirical comparison of two state-of-theart detection architectures YOLO-v3 and SSD was conducted, and according to the results the SSD architecture was chosen to solve the given problem, which allows to achieve optimal performance (mAP accuracy, FPS, amount of RAM and disk space) on a mobile device ASUS ZenFone Go ZB450KL 8GB. Keywords: neural networks, object recognition, object detection. DOI: 10.25791/asu.1.2023.1409 Pp. 24-33.