Entropy pooling in convolutional neural networks

An image recognition process is to learn the features of the objects and to recognize these learned features on previously unseen images. Convolutional neural networks are also a very popular structure used in image recognition processes. Convolutional neural networks are a special class of artificial neural networks and are the process of extracting hierarchical features from the images and making them meaningful through the artificial neural network. This process is done through various layers. From these layers, important features are extracted with various filters in the convolution layer, the purpose of this layer is to extract the most meaningful features. Pooling layer is after the convolution layer and this layer has a key position in convolutional neural networks. The main task of the pooling layer is to significantly reduce the size of the feature maps created through the convolution layer. Although the known basic pooling methods are simple and easy to process, they can be of low accuracy. In this study, which we will do, an entropy pooling method has been developed based on the concept of entropy, whose success is based on information theory and which gives successful results in extracting meaningful information from signals and images, and its performance will be compared with traditional methods. Test results are shared in detail.