Source-linked AI summary
Flip-Rotate-Pooling Convolution and Split Dropout on Convolution Neural Networks for Image Classification
Fa Wu, Peijun Hu, Dexing Kong
TL;DR
CNNs must control overfitting while improving invariance to image rotations. The paper introduces Split Dropout plus rotate-pooling and flip-rotate-pooling convolutions, and reports improved ImageNet classification with faster convergence and little added cost.
Problem
CNNs face overfitting in large-parameter models, while standard pooling does not reliably provide rotation invariance.
Method
The paper replaces standard Dropout with Split Dropout and adds RPC or FRPC, which pool responses from rotated or flipped filters.
Results
1.18% top-1 accuracy increase on ImageNet 2012 is reported for FRPC with sDropout versus the original network, alongside faster convergence and improved rotation robustness.
Takeaways & Limitations
The proposed techniques improve CNN classification and rotation robustness while adding very little memory or computation cost.
Takeaways & Limitations
RPC is not applied to all filters because object orientations can aid classification, and the experiments found r = 50% better than r = 100%.
Abstract
from arXiv · showhide
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.
1. Introduction
The paper addresses two CNN limitations: overfitting during training and weak rotation invariance. It proposes Split Dropout and rotational convolutions, then evaluates them on ImageNet 2012.
- 1. Introduction: CNN pooling encodes translation invariance but does not reliably capture rotation invariance, motivating rotational convolution methods.Experiments cited in the paper report that CNN outputs are not invariant to rotation transformations.
- 1. Introduction: Split Dropout keeps both randomly formed unit subsets, training all weights while still breaking co-adaptation.Standard Dropout discards half the units during each forward pass, whereas sDropout feeds the other subset into a second shared-weight pathway.
- 1. Introduction: RPC and FRPC rotate or flip convolution filters before convolution and max-pooling to extract rotation- or flipping-invariant features.The techniques are introduced without changing the original CNN architecture.
- 1. Introduction: On ImageNet 2012, sDropout lowers testing error and accelerates convergence, while RPC and FRPC improve classification accuracy on rotated images.The methods add very little memory and computation cost.
2. Related Work
Related work covers CNN architecture and regularization advances, prior Dropout variants, and approaches for learning transformation-invariant image representations. The paper positions its methods as lightweight alternatives that operate within convolutional networks.
- 2. Related Work: CNN progress has combined deeper architectures, improved activations, data augmentation, and regularization such as Dropout.The paper identifies improving Dropout as one of its major contributions.
- 2. Related Work: Prior Dropout extensions modify sampling, weights, or training objectives, while sDropout retains the discarded units as inputs to an additional shared-weight pathway.The cited alternatives include standout, DropConnect, and fast dropout.
- 2. Related Work: Existing invariance methods include unsupervised feature learning, tiled convolutions, wavelet representations, and learned or fixed transformations for scale or rotation invariance.These approaches differ in architecture, pretraining, or whether filters are learned.
- 2. Related Work: Data augmentation adds transformed training examples, whereas this paper pools over systematically transformed filters and can restrict rotation processing to upper layers.The approach is related to pooling features over symmetry groups within a neural network.
3. Approach
The approach replaces standard Dropout with sDropout, which trains both randomly split subsets, and adds RPC/FRPC layers that use rotated or flipped filters with max-pooling to encode rotational invariance.
- 3.1. Split Dropout: sDropout feeds both randomly split subsets into shared-weight subnetworks instead of discarding one subset, preserving Dropout-style regularization while training all weights each iteration.The method averages the subnetworks’ costs and gradients, retaining the regularization effect of broken co-adaptation while enabling faster convergence.
- 3. Approach: The approach targets a limitation of CNN pooling: translation invariance is encoded, but pooling units do not adequately capture rotation invariance.RPC and FRPC replace convolution layers while leaving the remaining network stages unchanged.
- 3.2. Rotate-Pooling Convolution: RPC rotates shared-weight convolution filters through eight orientations, convolves each with the input, and max-pools the resulting feature maps into one output.RPC is applied to high-level convolution layers such as conv31–conv33 and does not add trainable parameters.
- 3.3. Flip-Rotate-Pooling Convolution: FRPC extends the rotational convolution design by adding flipped versions of selected filters, producing two channels that are max-pooled into one output.The proportion and orientation of flipped filters are selected experimentally rather than applying flipping to every filter.
4. Implementation Details
Experiments use an 8-layer Zeiler–Fergus convnet, modifying fully connected and convolutional layers for sDropout and rotational convolution. Training uses synchronized two-GPU data parallelism.
- Architecture: The experiments use an 8-layer convnet architecture, with sDropout in fc1 and fc2 and RPC or FRPC in conv31, conv32, and conv33.Rotational layers use filter rotation and flipping percentages of 100%, 50%, or 25%.
- Rotational convolution: Rotational convolution layers apply the same selected rotation and flipping percentages to every sample.
- Parallel implementation: The implementation uses data parallelism on convolution layers with two GPUs, synchronizing before the first fully connected layer and processing 64 images per GPU per iteration.
5. Experiments
Experiments on ImageNet 2012 evaluate sDropout, RPC, and FRPC using Zeiler–Fergus-based CNNs, measuring convergence, computational cost, rotation robustness, and classification error. sDropout converges faster with little extra time, rotational convolutions improve robustness to rotated images, and the combined model achieves the best reported error reduction.
- 5.3. CNNs performance comparison: 1.18% error-rate decline is achieved by ZPSRF versus ZP on ImageNet 2012 using 10-view top-1 evaluation.ZPSRF combines sDropout with flip-rotate-pooling convolution; the models use the Zeiler and Fergus architecture with PReLU.
- 5.1. sDropout v.s. Dropout: sDropout makes model Z converge faster than standard Dropout and lowers its validation top-1 error from 37.72% to 37.19%.With 10-view testing, ZS also lowers top-5 error from 15.81% to 15.69% while taking little extra computing time.
- 5.2. RPC & FRPC: On 50k ImageNet validation images rotated from 0° to 360° into 64 views, Figure 5 compares average top-1 accuracy curves for ZP, ZPS, and ZPSR.With r = 100%, all filters in conv31, conv32, and conv33 are rotated; ZPS outperforms ZP and ZPSR, especially around 180°.
- 5.2. RPC & FRPC: ZPSR consistently outperforms ZP and ZPS on rotated images, especially near 180°, while ZPS is slightly better than ZP.Accuracy generally drops as rotation increases, with peaks at 90°, 180°, and 270° attributed to rotational symmetry in some images.
- 5.2. RPC & FRPC: Across image examples, ZPSR maintains higher true-label probability over rotation angles than ZPS, particularly for images poorly handled by ZPS.The advantage varies by image type: ZPS is already robust at symmetric angles for some images, whereas ZPSR performs well across more angles in other cases.
- 5.3. CNNs performance comparison: The final comparison evaluates eight model variants on the ImageNet 2012 50k validation dataset using a 10-view error rate under shared CNN architecture and parameter settings.The variants are Z, ZS, ZSRF (quarter), ZP, ZPS, ZPSR(full), ZPSR(half), and ZPSRF (quarter).
6. Discussion and Conclusion
The paper concludes that rotational convolution improves CNN classification, with the 25% rotate-and-flip FRPC setting combined with sDropout performing best. sDropout also accelerates convergence and slightly improves testing accuracy while requiring little extra memory or computation.
- The 25% rotate-and-flip setting with FRPC and sDropout improves the model most.
- Rotational convolution techniques improve CNN classification performance.
- sDropout makes training converge faster and slightly increases testing accuracy.
- The proposed techniques provide improvements with little additional memory or computing time.