Simulating the Behaviour of Choquet-Like (pre) Aggregation Functions for Image Resizing in the Pooling Layer of Deep Learning Networks

Abstract

The data volume expansion has generated the need to develop efficient knowledge extraction techniques. Most problems that are processed by these techniques have complex information to be identified and use different machine learning methods, such as Convolutional and Deep Learning Network. These networks may use a variety of aggregation functions to resize images in the pooling layer. This paper presents a study of the application of aggregation functions based on the generalizations of the Choquet integral, namely, the novel Choquet-like (pre) aggregation functions, in image dimensional reduction, simulating the pooling layer of a Deep Learning Networks. This paper is the natural evolution of the initial study where only the standard Choquet integral was applied. We compare the behaviour of such functions with the usual ones used in the literature, namely, the maximum and the arithmetic mean. A quantitative evaluation is done over an image dataset by using different image quality measures to compare the results.