Search

The research aims to improve structural methods of image classification based on a description as a set of keypoint descriptors. The focus is on implementing a classifier training using a committee of Kohonen networks separately for each description in an etalon database. The training result is a fixed set of data centroids, which ensures high classification speed. The improvement comprises implementing independent training for each etalon, which increases the accuracy of approximating descriptions with a set of centroids and, in general, guarantees that the classification efficiency remains at a decent level. Calculating the cluster centroids for each class prevents the influence of descriptors from other classes. In addition, independent training is effective in cases when the powers of etalon descriptions differ. The classification speed of the proposed method based on training a committee of Kohonen networks, compared to the traditional linear search method, increases proportionally to the ratio of the description power and the number of generated centroids. The presented results of experimental modeling of the proposed methods comprise a database of images of coins. The test sample is formed as a set of images from the etalon database by applying geometric transformations of shift, scale, and rotation. The testing stage has shown a high level of classification accuracy after a proposed training and revealed a practical opportunity to choose a network structure and parameters capable of providing the required level of accuracy and speed criteria for classification for an applied task.

The aim of the research is to improve the effectiveness of image recognition methods according to the description in the form of a set of keypoint descriptors. The focus is on increasing the speed of analysis and processing of description data while maintaining the required level of classification efficiency. The class of the image is provided as a description of the etalon. It is proposed to transform the description by implementing a statistical system of features for non-intersecting data fragments. The developed method is based on the aggregation of data distribution values within the description, the basis of which is the bit representation of the descriptors. Statistical features are calculated as the frequency of occurrence of the fixed value of a fragment on a set of description data and thus reflect the individual properties of images. Three main classifier models are analyzed: calculating the measure of data relevance in the form of distributions; assigning each of the descriptors to defined classes (voting); using the apparatus of statistical data analysis to decide on the significance of the difference between the distributions of the object and etalons. The results of software modeling of methods and calculations of statistical significance of differences based on distributions for training sets of images are represented. Using distributions instead of a set of descriptors increases the processing speed by hundreds of times, while the classification accuracy is maintained at a sufficient level and does not deteriorate compared to traditional voting.

The key task of computer vision is the recognition of visual objects in the analysed image. This paper proposes a method of searching for objects in an image, based on the identification of a cluster representation of the query descriptions and the current image of the window with the calculation of the relevance measure. The implementation of a cluster representation significantly increases the speed of identification or classification of visual objects while maintaining a sufficient level of accuracy. Based on the development of models for the analysis and processing of a set of descriptors of keypoints, we have obtained an effective method for the identification of visual objects. A comparative experiment with the traditional method has been conducted, where a linear search for the nearest descriptor was implemented for identification without using a cluster representation of the description. In the experiment, a speed gain for the developed method has been obtained in comparison with the traditional one by approximately 5.2 times with the same level of accuracy. The method can be used in applied tasks where the time of object identification is critical. The developed method can be applied to search for several objects of different classes. The effectiveness of the method can be increased by varying the values of its parameters and adapting to the characteristics of the data.

The subject of research of the article is the methods of image classification according to the set of descriptors of key points in computer vision systems. The aim is to increase the efficiency of classification by introducing a multicomponent data model on a set of descriptors for the base of reference images. Applied methods: ORB detector and descriptors, apparatus of set theory and vector space, metric models for determining the relevance of sets of multidimensional vectors, elements of probability theory, software modeling. Results are obtained: a modified method of image classification based on the introduction of a multicomponent model for data analysis with a system of centers is developed, methods of constructing a set of data centers are identified, the most effective is the set medoid and centers based on it. The effectiveness of the modification significantly depends on the method of forming the centers, the applied classification model, as well as on the data itself. The best results were shown by the classification with the integrated indicator separately for each of the standards in the form of the sum of the values of the distributions for the set of centers; experimentally tested the effectiveness of the classification, confirmed the efficiency of the proposed method. The practical significance of the work is the construction of classification models in the transformed data space, confirmation of the efficiency of the proposed modifications on the examples of images, the creation of software for the implementation of developed classification methods in computer vision systems.

The article proposes a new classification method based on implementing the high-speed search tools for the indexed data structure created on the etalon set of features, which has significant advantages in processing speed compared to the traditional approaches. The classifier is represented as two-stage processing, where at the first stage the class for the separate object descriptor is determined, and at the second stage, the resulting class of the object is determined based on the obtained set of local solutions. The developed method is based on the preliminary construction of the indexed hash structures for the set of descriptors of the base of the etalon images. Implementing the hash representation allows for increasing the speed of identification or classification of visual objects. A comparative experiment with the traditional method of voting has been conducted, where the linear search for the nearest descriptor has been implemented for the identification without the use of prior creation of the indexed hash representation of the etalons. In the experiment, we have gained in processing speed for the developed method compared to the traditional over 10 times. The gain in processing speed increases proportionally with the number of the etalons and the number of the descriptors in the descriptions. The experiment has shown that the efficiency of the method can be enhanced by varying the values of its parameters and adapting to the properties of the data.

The processes of intelligent data processing in computer vision systems have been researched. The problem of structural image recognition is relevant. This is a promising way to assess the degree of similarity of objects. This approach provides the simplicity of construction and the high reliability of decision making. The main problem of an effective description of characteristic features is the distortion of fragments of analyzed objects. The reasons for changing the input data can be the actions of geometric transformations, the influence of background or interference. The elements of false objects with similar characteristics are formed. The problem of ensuring high-quality recognition requires the implementation of effective means of image processing. Methods of statistical modeling, granulation of data and fuzzy sets, detection and comparison of keypoints on the image, classification and clustering of data, and simulation modelling are used in this research. The implementation of the proposed approaches provides the formation of a concise description of features or a vector representation of unique keypoints. The verification of theoretical foundations and evaluation of the effectiveness of the proposed data processing methods for real image bases is performed using the OpenCV library. The applied significance of the work is substantiated according to the criterion of data processing time without reducing the characteristics of reliability and interference immunity. The developed methods allow to increase the structural recognition of images by several times. Perspectives of research may involve identifying the optimal number of keypoints of the base set.

The article considers the problem of image recognition in computer vision systems. The results of the development of the method for image classification, using a structural approach, are presented. The classification method is based on calculating the values of statistical distributions for the set of description descriptors. The distribution vector for a fixed set of classes is based on the calculation of the degree of similarity with the integral characteristics for the descriptions of the etalon base. Two options for constructing the classifier on the principles of object – etalon and object descriptor – etalon, which differ in the degree of integration of the solution, are proposed. The median for the set of vectors describing the etalon is used as the aggregate characteristic of the etalon descriptions. The experimental evaluation of the effectiveness of the developed classifiers in terms of verification of performance and evaluation of the probability of correct classification according to the results of processing of applied images based on three etalons are carried out. The values of precision and completeness indicators for the method object descriptor – etalon, which has demonstrated the significant advantage over the integrated approach, are given. At the same time, both proposed in the experiment methods classify the set of etalons without error. The methods of mathematical statistics, intellectual data analysis, image recognition, the apparatus for calculating the relevance of the system of the features, as well as simulation modelling, are used in this research. Based on the study and the experiment, it was found that the processing time of the images for the developed method is approximately 7 times less than for the traditional method, without reducing the accuracy. The perspective of further research is to study the interference immunity of the developed methods and evaluate their applied effectiveness for three-dimensional image collections.

Предметом досліджень статті є класифікатори зображень за множиною дескрипторів ключових точок. Метою є підвищення продуктивності методів класифікації, зокрема, скорочення обчислювальних затрат шляхом впровадження на попередньому етапі оброблення апарату редукції для подання еталонних даних. Методи, що застосовуються: метричний апарат у векторному просторі, моделі для оцінювання інформативності даних, методи пошуку в масивах даних, моделі для визначення релевантності векторів та множин векторів, програмне моделювання. Отримані результати: розроблено метод редукції даних для задач класифікації зображень на основі впровадження метричних критеріїв для оцінювання інформативності елементів структурного опису зображення, що скорочує опис та прискорює обчислення; час класифікації для розглянутих експериментальних описів пропорційно скорочується зі зменшенням об’єму опису; для модельного експерименту досягнуто скорочення часу класифікації у п’ять разів при зменшенні обсягу опису у два рази; проведене моделювання підтверджує працездатність та результативність запропонованого методу в аспекті забезпечення якості класифікації з використанням засобів редукції. Практична значущість роботи – побудова моделей для оцінювання ступеня інформативності для образів візуальних даних; підтвердження працездатності запропонованих модифікацій засобів аналізу даних, розроблення прикладних програмних моделей для впровадження запропонованих методів редукції даних та класифікації зображень у системах комп’ютерного зору.

In this paper, we propose an investigation of the properties of structural image recognition methods in the cluster space of characteristic features. Recognition, which is based on key point descriptors like SIFT (Scale-invariant Feature Transform), SURF (Speeded Up Robust Features), ORB (Oriented FAST and Rotated BRIEF), etc., often relating to the search for corresponding descriptor values between an input image and all etalon images, which require many operations and time. Recognition of the previously quantized (clustered) sets of descriptor features is described. Clustering is performed across the complete set of etalon image descriptors and followed by screening, which allows for representation of each etalon image in vector form as a distribution of clusters. Due to such representations, the number of computation and comparison procedures, which are the core of the recognition process, might be reduced tens of times. Respectively, the preprocessing stage takes additional time for clustering. The implementation of the proposed approach was tested on the Leeds Butterfly dataset. The dependence of cluster amount on recognition performance and processing time was investigated. It was proven that recognition may be performed up to nine times faster with only a moderate decrease in quality recognition compared to searching for correspondences between all existing descriptors in etalon images and input one without quantization.