Recognition of anomalies on multitime panoramas using the neural network method of model amalgamation

A method for classifying anomalies on multitime panoramas using the neural network method of model amalgamation is described. The main idea of this method is to break the deep learning model into its component parts. Then, existing design patterns are selected and adapted for each part. To solve the problem of classifying anomalies, a composite template was developed and a new neural network model was designed. An improvement of the design pattern of neural networks ConvNet based on the technology of using “dense convolutional blocks” has been made. To classify anomalies, a new template has been developed — SOTA-ConvNet which includes a ready-made neural network model architecture suitable for working with panoramic images. The neural network model built on the basis of the SOTA-ConvNet template showed advantages over the neural network based on ConvNet: the accuracy of image classification increased and the number of trainable parameters decreased by an order of magnitude.