SYSTEM OF IMAGE RECOGNITION OF SURROUNDING OBJECTS WITH NEURAL NETWORKS
Kateryna O. Kirei · Collection of Scientific Publications NUS · 2021
The article considers one of the directions of automation of recognition of real world objects using the convolutional neural networks.The recognition of objects in an image and their classification is the basis of machine vision systems.However, real-world computer recognition systems do not always work well.The main problem is that in most cases it is not possible to properly identify the features on which recognition should be based.For tasks where such features can be identified, artificial recognition systems have become widespread and widely used.For example, recognition of vehicle license plates, human faces, etc.However, sufficient adequacy of selection and recognition of various objects of the real world in the time close to the real one is still not achieved, which does not allow to achieve the necessary indicators in real tasks.Purpose of research is the development of an efficient neural network algorithm capable of quickly and accurately recognizing real-world objects containing images from a predetermined list of objects.Method.basised on the analysis of algorithms and neural network architecture, the most appropriate type of neural network was chosen; the algorithm of object recognition in the image was optimized; the most suitable configuration of hyperparameterizations of neural networks was empirically chosen, directly influences the speed and performance of the neural network in recognition.Results.Several models have been developed for neural networks, such as precise, precise2, and fast, with specific hyperparameter configurations.Then the neural network was trained on the basis of these configurations with the COCO dataset and the main functions of the created system were tested.Scientific novelty is the development of an advanced algorithm for the recognition of objects and their borders on the image.Practical importance.A system has been developed that can quickly and accurately recognize and mark objects in an image.The system can also work with any data set and is very flexible in setting.