METHODS OF DETECTING DISTORTIONS IN IMAGES IN COMPUTER VISION SYSTEMS

The article is devoted to solving the problem of detecting image distortions in computer vision systems. Distortions can significantly reduce the accuracy of computer vision algorithms, leading to errors in object recognition, false positives, or missing critical events. Traditional methods of distortion detection include the analysis of brightness histograms, frequency characteristics, and spatial gradients. However, most existing solutions either have insufficient accuracy or require significant computing resources, which limits their application. This paper considers a hybrid approach to distortion detection that combines statistical methods of image analysis and optical flow density calculation. This approach allows you to cover a large number of different types of distortions and effectively detect them with minimal computational costs, which is especially important for real-time systems. To develop the approach, methods for detecting artifact images were analyzed: brightness analysis, low contrast testing, vertical stripe detection, and block distortion detection. A comparative analysis of these methods showed that none of them allows you to effectively detect various distortions in images. To improve the efficiency of solving this problem, it is proposed to use a combined method that includes all the considered methods. The method sequentially checks the image for overexposed areas, low contrast, vertical stripes and block distortions, and then forms the final output. If any of the detectors detects significant distortions, the image is marked as artifactual. To assess the efficiency of the proposed method, it was tested on a sample of 700 images obtained from street surveillance cameras in the city of Kazan. Half of the images were without distortions, and the other half had various types of distortions and with varying degrees of their severity. Based on the testing results, 363 true positive classification results, 8 false positive results and 329 true negative results were identified. The following classification quality metrics were calculated: accuracy = 98.86%, precision = 97.84%, recall = 100% and F1-score = 98.91%. The obtained results allow us to judge the high efficiency of the combined method for detecting distortions in images. The results of the work can be successfully applied in monitoring systems using video surveillance and in other subject areas where high accuracy of image processing is critically important.