Development of a Neural Network Method in the Problem of Classification and Image Recognition

Abstract

In the operation of any algorithm for face recognition or face detection, two logical blocks should be distinguished: an extractor of characteristic features and a classification mechanism. The action of the extractor is based on the extraction of information useful for the classifier from a huge stream of input data. When identifying a person, this information may be the characteristics of uniquely determined features (for example, the relative position of the eyes, eyebrows, lips and nose used in forensic science). When deciding whether to assign a class label to a recognizable object, a classifier should be guided by these very features. Feature selection is the most important task. Obviously, when choosing them, the most unique properties are taken into account, since they are the most reliable way to judge whether an object belongs to a particular class. There are many different approaches to obtaining class traits. The application of Object Detection to the solution of the problem of image classification and recognition is considered. The description of the FastER-RCNN method based on a two-stage neural network is given. The results of applying the YOLOv3 algorithm for training a neural network with different steps are presented. It is proposed to use an improved approach based on YOLO for accurate and fast object detection. The contributions of this work are: an efficient and accurate real-time detection model, ease and ability to locate objects based on improvements to the Fast-RCNN algorithm.

Keywords

• face recognition
• image recognition
• convolutional neural networks
• regional convolutional neural networks (r-cnn) model
• bounding box
• anchor