Adaptive Pooling for Convolutional Neural Networks with Arbitrary Input Sizes
Hsi‐Chin Hsin, Chien-Kun Su · 2021 IEEE 3rd Eurasia Conference on IOT, Communication and Engineering (ECICE) · 2021
Convolutional neural networks have been widely used in deep learning recently. This paper presents an adaptive scheme to modify the input layers of the conventional convolutional neural networks such that images of arbitrary sizes can be directly input. Specifically, motivated by the advantage of content-aware image resizing, which takes the regions of interest into account for effective displaying on various screens with different dimensions and aspect ratios, it is beneficial to incorporate content-aware image resizing into convolutional neural networks. Experimental results show that image classification can be improved in terms of mean average precision.