The use of affine transformations for image correction with further use in recognition systems

Abstract

Today, neural networks are gaining popularity and are increasingly used in various fields of life. They are used for data analysis, clustering, classification, object detection, or hidden patterns. Neural networks are increasingly used in business forecasting tasks. Computer vision systems are no exception. An important component of these processes is image preprocessing. It is known that images obtained from video cameras can exhibit perspective distortions originating from imperfect camera angles. Usually, the affine transform technique is used to correct geometric distortions or deformations, such as scaling, moving, shifting, rotating. In this paper, we study the features of affine transforms and their application with neural networks. Methods for recognizing geometric parameters in image transformation and moving are considered. Attention is paid to elementary transformations, which include transfer, scaling, shift, rotation. A mathematical model of image transformation for computer vision systems is developed. On the basis of the proposed method, an algorithm for perspective transformations of images obtained from video cameras located in parking lots or city parking lots is built, which greatly facilitates the further detection, segmentation and classification of objects. To improve the performance of the classical Mask R-CNN, a study was conducted in which a block with affine transformations was added to the convolutional neural network. Affine transformations are used to correct the perspective convergence of lines in the frame that are parallel in reality.

Keywords

• video images
• geometric parameters
• computerized system