Flip-Rotate-Pooling Convolution and Split Dropout on Convolution Neural Networks for Image Classification

By Fan Wu, Peijun Hu, Dexing Kong. 2015

30 citations recorded by the source corpus.

This paper presents a new version of Dropout called Split Dropout (sDropout) and rotational convolution techniques to improve CNNs' performance on image classification. The widely used standard Dropout has advantage of preventing deep neural networks from overfitting by randomly dropping units during training. Our sDropout randomly splits the data into two subsets and keeps both rather than discards one subset. We also introduce two rotational convolution techniques, i.e. rotate-pooling convolution (RPC) and flip-rotate-pooling convolution (FRPC) to boost CNNs' performance on the robustness for rotation transformation. These two techniques encode rotation invariance into the network without adding extra parameters. Experimental evaluations on ImageNet2012 classification task demonstrate that sDropout not only enhances the performance but also converges faster. Additionally, RPC and FRPC make CNNs more robust for rotation transformations. Overall, FRPC together with sDropout bring $1.18\%$ (model of Zeiler and Fergus~\cite{zeiler2013visualizing}, 10-view, top-1) accuracy increase in ImageNet 2012 classification task compared to the original network.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Media Technology

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

1. Qiang Qiu
2. Sander Dieleman
3. Guillermo Sapiro
4. Xiuyuan Cheng
5. Deng Cai
6. Qixiang Ye
7. Devis Tuia
8. Jianbin Jiao

In Computer Vision and Pattern Recognition

1. Luc Van Gool
2. Dacheng Tao
3. Yi Yang
4. Ming-Hsuan Yang
5. Andrew Zisserman
6. Devis Tuia
7. Jitendra Malik
8. Xiaogang Wang

In Media Technology

1. Devis Tuia
2. Diego Marcos
3. Deng Cai
4. Robert Calderbank
5. Haifeng Liu
6. Qixiang Ye
7. Guillermo Sapiro
8. Xiuyuan Cheng

Citations in the corpus

• RotDCF: Decomposition of Convolutional Filters for Rotation-Equivariant Deep Networks (2018, methodology): “[8, 38] proposed rotating conventional filters to perform rotation-invariant texture and image classification.”
• Exploiting Cyclic Symmetry in Convolutional Neural Networks (2016, methodology): “Wu et al. (2015) apply rotated and flipped copies of each filter in a convolutional layer and then max-pool across the resulting activations.”
• Oriented Response Networks (2017, background): “…rotating conventional filters to perform rotation-invariant texture and image classification; however, without upgrading conventional filters to multi-oriented filters with orientation channels, their capability about capturing hierarchical and fine-detailed orientation information remains limited.”
• Learning rotation invariant convolutional filters for texture classification (2016, methodology): “In [20], authors applied this concept to the rotation group in the upper layers of a deep CNN in order to obtain invariance to global rotations of the input image.”
• Deep Rotation Equivariant Network (2017, methodology): “One widely used method to achieve rotation equivariance is to constrain the filters of the first convolutional layer to be rotated copies of each other, and then apply cross-channel pooling immediately after the first layer [7]–[9].”