Pooling Operations in Deep Learning: From “Invariable” to “Variable”

Abstract

Deep learning has become a research hotspot in multimedia, especially in the field of image processing. Pooling operation is an important operation in deep learning. Pooling operation can reduce the feature dimension, the number of parameters, the complexity of computation, and the complexity of time. With the development of deep learning models, pooling operation has made great progress. The main contributions of this paper on pooling operation are as follows: firstly, the steps of the pooling operation are summarized as the pooling domain, pooling kernel, step size, activation value, and response value. Secondly, the expression form of pooling operation is standardized. From the perspective of “invariable” to “variable,” this paper analyzes the pooling domain and pooling kernel in the pooling operation. Pooling operation can be classified into four categories: invariable of pooling domain, variable of pooling domain, variable of pooling kernel, and the pooling of invariable “+” variable. Finally, the four types of pooling operation are summarized and discussed with their advantages and disadvantages. There is great significance to the research of pooling operations and the iterative updating of deep learning models.

Deep Multi-Magnification Similarity Learning for Histopathological Image Classification

“…C under the magnification m, and where h and w represent the height and width of the image feature map. In order to calculate the different magnification features with different shapes, we changed different features into the target same shape by adaptive pooling [46].…”

Brain tumor feature extraction and edge enhancement algorithm based on U-Net network

“…The study showcased the model's impressive performance, achieving a top-1 accuracy of 87.8% and a mIoU measure of 56% on the ADE20K dataset. With the goal to reach global information capture, Zhou et al [8] conducted an assessment of four pooling approaches and devised two strategies, namely “variable” and “invariant”, to enhance the scope of sensory fields. Ding et al [9] incorporated the principles of ACNet into the underlying framework of VGG in the RepVGG design.…”