Visibility loss detection for video camera using deep convolutional neural networks

Abstract

The article describes the application of various machine learning methods for the analysis of images obtained from a video camera with the purpose of detection its partial or total visibility loss. Computational experiments were performed on a data set containing more than 6800 images. Support vector machine, categorical boosting and simplified modifications of VGG, ResNet, InceptionV3 architectures of neural networks are used for image classification. A comparison of the methods quality is presented. The best results in terms of classification accuracy are obtained using ResNetm and InceptionV3m architectures. The recognition accuracy is on the average more than 96%. The processing time per reduced input frame is 8–12 ms. The obtained results confirm the applicability of the proposed approach to the detection of camera visibility loss for real tasks arising in on-board machine vision systems and video surveillance systems.