Convolutional neural networks for solving computer vision problems

This article provides an overview of the main methods of solving computer vision problems of classification, segmentation and image processing, which are implemented in CV systems. Computer vision systems are programmed to perform highly specialized tasks, capable of detecting objects during identification, reading serial numbers, and searching for surface defects. When applying deep learning methods in CV systems, their processing speed on large data sets and the accuracy of image classification/segmentation are significantly increased. Artificial vision systems are able to identify individual pixels according to the relevant features during processing, provide a high-quality result in pattern recognition, image restoration, and fitting part of the image. Although some computer vision algorithms were developed to simulate visual perception, a larger number of proposed methods are able to fully process images and determine their characteristic properties. The scope of application of CV systems will continue to expand, as the need for artificial intelligence systems is growing rapidly. The purpose of this article is to provide a structured review of computer vision technologies based on their advantages and disadvantages. The work summarizes the types of CV-systems with artificial intelligence according to the spectrum of their applications, highlights the main problematic areas of their research, such as recognition, identification and detection. The article reviews convolutional neural networks (CNNs), which are successfully applied to the analysis of visual images in deep learning. CNN architectures in some cases outperform artificial neural networks in classification tasks by their performance. Currently, convolutional neural networks are the main tool for classification and recognition of objects, faces in photographs, recognition of video and audio materials. This paper provides a comparative analysis of well-known CNN models: LeNet 5, AlexNet, VGGNet, GoogLeNet, ResNet and their effectiveness in CV systems. Approaches to the modeling of architectures of convolutional neural networks are proposed, which will allow, in the future, to solve the problem of classification in tasks for computer vision, thereby increasing their performance, accuracy and quality of processing.