МОДЕЛЬ І МЕТОД НАВЧАННЯ КЛАСИФІКАТОРА КОНТЕКСТІВ СПОСТЕРЕЖЕННЯ НА ЗОБРАЖЕННЯХ ВІДЕОІНСПЕКЦІЇ СТІЧНИХ ТРУБ

Abstract

Video inspection is often used to diagnose sewer pipe defects. To correctly encode founded defects according to existing standards, it is necessary to consider a lot of contextual information about the orientation and location of the camera from sewer pipe video inspection. A model for the classification of context on frames during observations in the video inspection of sewer pipes and a five-stage method of machine learning is proposed. The main idea of the proposed approach is to combine the methods of deep machine learning with the principles of information maximization and coding with self-correcting Hamming codes. The proposed model consists of a deep convolutional neural network with a sigmoid layer followed by the rounding output layer and information-extreme decision rules. The first stages of the method are data augmentation and training of the feature extractor in the Siamese model with softmax triplet loss function. The next steps involve calculating a binary code for each class of recognition that is used as a label in learning with a binary cross-entropy loss function to increase the compactness of the distribution of each class's observations in the Hamming binary space. At the last stage of the training method, it is supposed to optimize the parameters of radial-basis decision rules in the Hamming space for each class according to the existing information-extreme criterion. The information criterion, expressed as a logarithmic function of the accuracy characteristics of the decision rules, provides the maximum generalization and reliability of the model under the most difficult conditions in the statistical sense. The effectiveness of this approach was tested on data provided by Ace Pipe Cleaning (Kansas City, USA) and MPWiK (Wroclaw, Poland) by comparing learning results according to the proposed and traditional models and training schemes. The obtained model of the image frame classifier provides acceptable for practical use classification accuracy on the test sample, which is 96.8 % and exceeds the result of the traditional scheme of training with the softmax output layer by 6.8 %.