Improving the quality of video processing based on the use of an artificial YOLO neural network
V. E. Dementyev, А. Г. Ташлинский, Marat Suetin, Maria Gaponova · 2022 VIII International Conference on Information Technology and Nanotechnology (ITNT) · 2022
The paper considers the processing of video images of metal structures that potentially contain defects. Such processing is proposed to be performed with the use of convolutional neural networks like Yolo. To improve the quality of processing, options are offered in using the results of processing previous frames of a video sequence. In this case, it is supposed to make a prediction of the placement of previously selected objects on subsequent frames using pseudo-gradient procedures to estimate deformation parameters in a video sequence. For the areas of possible location of defects predicted in this way, it is proposed to perform an adaptive adjustment of the neural network hyperparameters that determine the probability of a false positive. The conclusion has been drawn about the viability and effectiveness of such approaches.