Revisiting Data Augmentation for Rotational Invariance in Convolutional Neural Networks

Quiroga, Facundo; Ronchetti, Franco; Lanzarini, Laura Cristina; Bariviera, Aurelio F.

doi:10.1007/978-3-030-15413-4_10

Cited by 20 publications

(18 citation statements)

References 9 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Measuring invariance to transformations was also tackled from an adversarial perspective [5], confirming that simple rotations or translations can have a big impact on performance. In [16,11,19], the effect of using different data augmenta tion schemes and CNNs architectures was measured and compared. Specifically, the translation sensitivity map developed in [11] related the classifier accuracy with the center position of the object in the image.…”

Section: Related Workmentioning

confidence: 99%

“…Typical C-NNs rely on feed-forward architectures with a series of convolu tional layers followed by one or two dense layers. These models, commonly trained with stochastic gradient descent and without data augmentation, can not learn invariances or equivariances to rotations [16,1]. Feed-forward networks exclusively composed of dense layers can approximate smooth functions given enough parameters.…”

Section: Introductionmentioning

confidence: 99%

“…Feed-forward networks exclusively composed of dense layers can approximate smooth functions given enough parameters. They can even learn arbitrary invariance and equivariance properties with heavy data augmentation [16]. Besides, some CNN models avoid dense layers [1 !…”

Section: Introductionmentioning

confidence: 99%

“…While applied transformations are often mild, full rotation invariance is possible by including all transformation angles or deformations. This approach was stud ied for Deep Restricted Boltzmann Machines [11], HOGs and C-NNs [15,16,19]. Although employed architectures are simpler, they generally require more train ing epochs to explore the wide space of rotated inputs.…”

Section: Introductionmentioning

confidence: 99%

“…Many proposed layers are individually invariant or equivariant, but no analysis was reported to under stand how networks as a whole encode such properties. Several authors studied which methods work best to achieve invariance [16,19], but no guiding principle in their analysis was employed except comparing the output accuracy. To the best of our knowledge, no works measure the internal invariance or equivariance of the network.…”

Section: Introductionmentioning

confidence: 99%

See 4 more Smart Citations

Measuring (in)variances in Convolutional Networks

Quiroga

Torrents-Barrena

Lanzarini

et al. 2019

Communications in Computer and Information Science

Self Cite

View full text Add to dashboard Cite

Convolutional neural networks (CNN) offer state-of-the-art performance in various computer vision tasks such as activity recogni tion, face detection, medical image analysis, among others. Many of those tasks need invariance to image transformations (i.e.. rotations, transla tions or scaling). This work proposes a versatile, straightforward and interpretable mea sure to quantify the (in)variance of CNN activations with respect to transformations of the input. Intermediate output values of feature maps and fully connected layers are also analyzed with respect to different input transformations. The technique is applicable to any type of neu ral network and/or transformation. Our technique is validated on ro tation transformations and compared with the relative (in)variance of several networks. More specifically, ResNet, AllConvolutional and VGG architectures were trained on CIFAR10 and MNIST databases with and without rotational data augmentation. Experiments reveal that rotation (in)variance of CNN outputs is class conditional. A distribution analysis also shows that lower layers are the most invariant, which seems to go against previous guidelines that recommend placing invariances near the network output and equivariances near the input.

show abstract

Section: Related Workmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

Measuring (in)variances in Convolutional Networks

Quiroga

Torrents-Barrena

Lanzarini

et al. 2019

Communications in Computer and Information Science

Self Cite

View full text Add to dashboard Cite

show abstract

Novel convolutional neural networks for efficient classification of rotated and scaled images

Tarasiuk

Szczepaniak

2021

Neural Comput & Applic

View full text Add to dashboard Cite

This paper presents a novel method for improving the invariance of convolutional neural networks (CNNs) to selected geometric transformations in order to obtain more efficient image classifiers. A common strategy employed to achieve this aim is to train the network using data augmentation. Such a method alone, however, increases the complexity of the neural network model, as any change in the rotation or size of the input image results in the activation of different CNN feature maps. This problem can be resolved by the proposed novel convolutional neural network models with geometric transformations embedded into the network architecture. The evaluation of the proposed CNN model is performed on the image classification task with the use of diverse representative data sets. The CNN models with embedded geometric transformations are compared to those without the transformations, using different data augmentation setups. As the compared approaches use the same amount of memory to store the parameters, the improved classification score means that the proposed architecture is more optimal.

show abstract