Methods for real-time object searching and recognizing in video images on ios mobile platform

The features of the most common methods and systems for searching and recognizing objects in video are explored. The research shows the feasibility of building search and recognition tools for the iOS platform in real time. The method of functional adaptation of the algorithm of search and recognition of objects to features of video is offered, which consists in processing of video image by smoothing and minimization filters, which reduces the time of search and recognition of objects. The block diagram and algorithm of system functioning were designed. Developed a program to solve the problem of finding and quickly recognizing objects in real time in Swift language on the iOS mobile platform. A convolutional neural network with YOLOv3 architecture was used along with framework for working with neural networks for mobile CoreML applications. A method of improving the performance of such a neural network is proposed, which is based on the quantization of the neural network weights and minimizes the model size and search time of its objects. The frequencies of image processing using the proposed means and models of neural networks of the type YOLOv3-tiny, YOLOv3-416 and our own model YOLOv3-KD are investigated. The possibility of functioning of the proposed funds in real time is provided.