Choice of Neural Network Architecture when Recognizing Objects that do not Have High-Level Features
This article explores the capabilities of pretrained convolutional neural networks in relation to the problem of recognizing defects for which it is impossible to identify any abstract features. The results of training the convolutional neural network AlexNet and the fully connected classifier of the VGG16 network are compared. The efficiency of using a pretrained neural network in the problem of defect recognition is demonstrated. A graph of the change in the proportion of correctly recognized images in the process of training a fully connected classifier is presented. The article attempts to explain the efficiency of a fully connected neural network classifier trained on a critically small training dataset with images of defects. The work of a convolutional neural network with a fully connected classifier is investigated. The classifier allows for classification into five categories: «crack» type defects, «chip» type defects, «hole» type defects, «multi hole» type defects and «defect-free surface». The article provides examples of convolutional network activation channels, visualized for each of the five categories. The signs of defects on which the activation of the network channels takes place are formulated. The classification errors made by the network are analyzed. The article provides predictive probabilities, below which the result of the network operation can be considered doubtful. Practical recommendations for using the trained network are given.