Automatic tuning of the pooling operation in convolutional neural networks for coal rock classification

Convolutional neural networks are used to automate image processing tasks such as classification, segmentation, object detection, style transfer, etc. These networks are actively applied in coal mining industry for automatic classification of coal rocks with high accuracy based on raw images. Accurate classification of the coal rocks is important for coal quality assessment, optimization of coal mining, preparation and processing. The main mathematical operations of convolutional networks are convolution and pooling. The paper discusses a generalization of the pooling operation. Usually the type of pooling is specified in advance by some aggregation operation, i.e. the average pooling or max pooling. The size of the aggregated area is also specified in advance. The type of pooling and the size of the aggregated region significantly affect the quality of coal-bed image processing. This paper proposes several parametric generalizations of the pooling operation, which cover the average and max pooling as special cases. A parametric generalization is also proposed for max pooling, which allows to vary the size of the aggregation region. Parameters of the proposed pooling generalizations are automatically trained along with the rest of the network weights.