RESEARCH OF THE PROCESS OF CREATING CLASS MARKERS IN IMAGES TO INCREASE THE RELIABILITY OF THE WORK OF CONVERTIBLE NEURAL NETWORKS

Abstract

Convolutional neural networks (CNNs) are a fundamental tool in the field of artificial intelligence for visual data analysis. Their distinguishing feature is the way they process images, which mimics the vision mechanisms of living organisms, using a multi-level approach to extracting features of objects in an image.

The concept of convolution, a special operation that applies filters or kernels to the input image to highlight important features, such as edges, corners, and textures, is at the heart of SNM. This allows the SNM to focus on key details, ignoring minor information, which simplifies further processing.

After convolution, the resulting information passes through a subsampling layer, where the image is reduced in size to reduce computational complexity and to increase invariance to changes in the input, such as scaling or rotation.

This process is repeated in several layers, where each subsequent layer reveals increasingly complex features based on those identified by previous layers. This hierarchical definition of features allows the network to "understand" complex visual patterns.

At the end of the process, after multi-layer convolutions and subsampling, the obtained characteristics are fed into one or more fully connected layers, where all the studied features are combined to solve the final task - whether it is image classification, object presence detection, or recognition.

ANNs are trained using an error backpropagation method, where the weights of each filter are adjusted based on the discrepancy between the predicted and actual results, thus gradually increasing the accuracy of the network.

ANNs have proven effective in a wide range of applications, from automatic license plate recognition to medical image diagnostics, thanks to their ability to separate important information from noise and process complex visual data.

The article describes the solution to the problem of preliminary image processing for the next stage of recognition and classification using convolutional neural networks. The identification of characteristic features of classes has been thoroughly investigated. The scientific novelty of the article is the proposal to transform the image into a state with the highlighting of high-priority features and the removal of low-priority features, followed by further processing.