CASPIAN JOURNAL
MANAGEMENT AND HIGH TECHNOLOGIES
ANALYSIS OF COMPUTER VISION METHODS FOR DETECTING PROHIBITED SYMBOLS ON IMAGES ON THE INTERNET
 |
Read | Shevchenko Viktor D., Marienkov Alexander N., Khanova Anna A. ANALYSIS OF COMPUTER VISION METHODS FOR DETECTING PROHIBITED SYMBOLS ON IMAGES ON THE INTERNET // Caspian journal : management and high technologies. — 2022. — №2. — pp. 9-18. |
Shevchenko Viktor D. - Astrakhan State University, Astrakhan, Russian Federation
Marienkov Alexander N. - Astrakhan State University, Astrakhan, Russian Federation
Khanova Anna A. - Astrakhan State University, Astrakhan, Russian Federation
Structural schemes and mathematical support of computer vision methods are considered: object search by template, binary classification using a convolutional neural network, object detection using a convolutional neural network with various architectural solutions. The implementation of all these methods is adapted to detect prohibited symbols on Internet images, in case of its presence and the announcement that the image is prohibited. As well as the recognition of symbols that are similar in structure to prohibited ones, but are not such, and the announcement that the image is not prohibited. Based on the set goal and formulated tasks, three computer vision methods were tested: object search by template, binary classification using a convolutional neural network, object detection using a convolutional neural network based on the YOLOv3 model. A test dataset was created that included 40 images, which was used to determine the accuracy of each method. The efficiency results for each considered method are obtained. Based on the task, three computer vision methods were analyzed: object search by template, binary classification using a convolutional neural network, object detection using a convolutional neural network. The method of object detection using the convolutional neural network of the YOLOv3 model proved to be the best. The percentage of accuracy of this method was 95%.
Key words: computer vision, forbidden images, convolutional neural network, binary classification, pattern search, object detection, YOLOv3