An Automatic Method for Interest Point Detection

• Main Author: I. G. Zubov
• Format: Article
• Language: Russian
• Published: Saint Petersburg Electrotechnical University "LETI" 2020-12-01
• Series: Известия высших учебных заведений России: Радиоэлектроника
• Subjects: convolutional neural networks; activation map analysis; interest point detection
• Online Access: https://re.eltech.ru/jour/article/view/474

Introduction. Computer vision systems are finding widespread application in various life domains. Monocularcamera based systems can be used to solve a wide range of problems. The availability of digital cameras and large sets of annotated data, as well as the power of modern computing technologies, render monocular image analysis a dynamically developing direction in the field of machine vision. In order for any computer vision system to describe objects and predict their actions in the physical space of a scene, the image under analysis should be interpreted from the standpoint of the basic 3D scene. This can be achieved by analysing a rigid object as a set of mutually arranged parts, which represents a powerful framework for reasoning about physical interaction.Objective. Development of an automatic method for detecting interest points of an object in an image.Materials and methods. An automatic method for identifying interest points of vehicles, such as license plates, in an image is proposed. This method allows localization of interest points by analysing the inner layers of convolutional neural networks trained for the classification of images and detection of objects in an image. The proposed method allows identification of interest points without incurring additional costs of data annotation and training.Results. The conducted experiments confirmed the correctness of the proposed method in identifying interest points. Thus, the accuracy of identifying a point on a license plate achieved 97%.Conclusion. A new method for detecting interest points of an object by analysing the inner layers of convolutional neural networks is proposed. This method provides an accuracy similar to or exceeding that of other modern methods.