Making a More Reliable Classifier via Random Crop Pooling

Abstract

Deep neural networks have achieved state-of-the-art performance for a variety of pattern-recognition tasks. In particular, the deep convolutional neural network (CNN), which is composed of several convolutional layers with a nonlinear activation function, pooling layers, and fully connected layers or an optional global average pooling layer, has received significant attention and is widely used in computer vision. Some research is now replacing a top fully connected layer with global pooling to avoid overfitting in the fully connected layers and to achieve regularization. This replacement is very important because global pooling with additional convolutional layers can eliminate restrictions on the necessity for fixed-size or fixed-length input in the fully connected layers. In this paper, the top global pooling layer is focused on, which is used in place of the fully connected layer and creates a simple and effective pooling operation called random crop (RC) pooling. Additionally, how to attain regularization in the top RC pooling layer is discussed. RC pooling randomly crops the feature maps so that only the images with sufficiently scaled and centered objects can be well-trained. This approach achieves comparable accuracy on the CIFAR-10/100 and MNIST.