Source-linked AI summary
Improving Deep Learning using Generic Data Augmentation
Luke Taylor, Geoff Nitschke
TL;DR
Limited and costly-to-collect training data can lead CNNs to over-fit, motivating data augmentation with label-preserving transformations. This study benchmarks generic geometric and photometric schemes on coarse-grained Caltech101 using a simple CNN and 4-fold cross-validation. Geometric methods generally outperform photometric methods, with cropping producing the largest reported Top-1 improvement of 13.82%.
Problem
Small training data-sets make CNNs prone to over-fitting, while comprehensive comparisons of popular augmentation methods on large coarse-grained data-sets were lacking.
Method
The study benchmarks popular geometric and photometric augmentation schemes on coarse-grained Caltech101 using a relatively simple CNN and cross-validation accuracy evaluation.
Results
Geometric augmentation schemes outperformed photometric schemes, and cropping produced a 13.82% Top-1 classification accuracy improvement.
Takeaways & Limitations
For coarse-grained data-sets, transformations altering image geometry were more effective than transformations altering lighting and color.
Takeaways & Limitations
Future work is needed to test whether the Caltech101 results transfer to other data-sets and CNN architectures, augmentation methods, and combinations.
Abstract
from arXiv · showhide
Deep artificial neural networks require a large corpus of training data in order to effectively learn, where collection of such training data is often expensive and laborious. Data augmentation overcomes this issue by artificially inflating the training set with label preserving transformations. Recently there has been extensive use of generic data augmentation to improve Convolutional Neural Network (CNN) task performance. This study benchmarks various popular data augmentation schemes to allow researchers to make informed decisions as to which training methods are most appropriate for their data sets. Various geometric and photometric schemes are evaluated on a coarse-grained data set using a relatively simple CNN. Experimental results, run using 4-fold cross-validation and reported in terms of Top-1 and Top-5 accuracy, indicate that cropping in geometric augmentation significantly increases CNN task performance.
1 Introduction
Limited training data can make CNNs prone to over-fitting, motivating data augmentation and comparisons of generic augmentation methods. The study evaluates geometric and photometric schemes on coarse-grained data to guide method selection.
- Small data-sets make CNNs prone to over-fitting, inhibiting generalization to unseen invariant data.
- Data Augmentation artificially inflates data-sets with label preserving transformations that add more invariant examples.
- Geometric transformations alter image geometry, while photometric transformations modify color channels to target invariance to position, orientation, lighting, or color.
- Complex DA uses domain-specific synthesization to produce richer training data, but generic DA is computationally inexpensive and easy to implement.
- Prior studies reported benefits from cropping, occlusion, flipping, and rotating on fine-grained or coarse-grained classification data-sets.
- The study comparatively evaluates popular geometric and photometric augmentation schemes on the coarse-grained Caltech101 data-set using a relatively simple CNN.
2 Data Augmentation (DA) Methods
The study defines data augmentation as label-preserving transformations that expand the training set, evaluating seven geometric and photometric methods against no augmentation.
- Data Augmentation Definition: Data augmentation maps original samples to label-preserving variants and forms an expanded set S′ = S ∪ T.If x belongs to class y, its transformed version φ(x) remains in class y.
- Methods Evaluated: Seven augmentation methods were evaluated: no augmentation, three geometric methods, and three photometric methods.No augmentation served as the benchmark for all experiments.
- Geometric Methods: Geometric augmentation preserves image shape while changing pixel destinations, position, or orientation through flipping, rotation, and cropping.Rotation experiments used angles of −30° and +30°; cropping extracted 224 × 224 regions from four corners and the center of 256 × 256 images.
- Geometric Methods: Flipping mirrors images across the vertical axis and is computationally efficient because it reverses image-matrix rows.The method requires only row reversal in the image matrices.
- Photometric Methods: Photometric augmentation shifts RGB values without changing geometry, using color jittering, edge enhancement, and fancy PCA.Fancy PCA adds multiples of principal components to training images and was performed on individual images because of computation and memory constraints.
- Photometric Methods: Edge enhancement intensifies image contours using Sobel edge detection, then composites the transformed result with the source image.The method was introduced as a new augmentation scheme based on the CNN’s ability to identify shapes.
3 CNN Architecture
The paper uses a relatively simple CNN designed to balance task performance with training speed while providing sufficient capacity to fit the data.
- Architecture Design: The CNN was designed with a favorable tradeoff between task performance and training speed for seven trainings using 4-fold cross-validation.Training datasets ranged from approximately 8.5 to 42.5 thousand images.
- Network Topology: The architecture contains five trainable layers: three convolutional layers, one fully connected layer, and one softmax layer.Its input is a three-channel RGB image, and the final softmax layer has size 101.
- Design Choices: Overlapping pooling was used to reduce over-fitting, while a 140-neuron fully connected layer was selected because larger sizes did not improve performance.Exploratory experiments found that smaller layer sizes produced richer encodings and better generalization.
- Training Configuration: The CNN used ReLU activations, Xavier initialization, Nesterov optimization with 0.90 momentum, and L2 regularization of 5 · 10^-4.The learning rate was 0.01, and the settings were based partly on related work and partly on exploratory experiments.
4 Experiments
Experiments benchmarked augmentation methods on Caltech101 using standardized images, 4-fold cross-validation, and repeated CNN training.
- Dataset and Validation: Caltech101 provided 8,421 evaluated images from 102 categories after exclusions and class-size trimming.The original dataset contained 9,144 images; 4-fold cross-validation used three subsets for training and one for validation.
- Preprocessing: Images were resized within 256 × 256 black canvases so every augmentation scheme operated on full images at a fixed resolution.Pixel values were then normalized from [0, 255] to [0, 1].
- Training Procedure: Each CNN was trained for 30 epochs, with implementation in Java 8 using DL4j and experiments run on an NVIDIA Tesla K80 GPU.The epoch count was selected through exploratory evaluation of validation and test scores.
5 Results and Discussion
Across the evaluated schemes, data augmentation improved CNN classification, with geometric methods generally outperforming photometric methods. Cropping produced the largest Top-1 improvement, while the mechanisms behind geometric–photometric differences remain unresolved.
- All data-augmentation schemes increased CNN classification performance, with standard deviations of 0.5% ∼1% across cross-validation folds.The exception was flipping, whose Top-5 score was inferior to Fancy PCA’s Top-5 score.
- Geometric augmentation generally outperformed photometric augmentation on both Top-1 and Top-5 scores.Flipping was the exception for Top-5 performance, where Fancy PCA scored higher.
- 13.82% was the cropping scheme’s Top-1 classification-accuracy improvement, the largest reported gain.Cropping represents specific translations and yielded a similar improvement for Top-5 classification.
- The study hypothesizes that cropping helps by generating more samples and exposing the CNN to a greater receptive view of training images.The authors connect the additional samples to reduced over-fitting and improved generalization, while describing this explanation as a hypothesis.
- Photometric methods produced smaller gains, including 1.44% Top-1 improvement for color jittering and 3.04% Top-5 improvement for Fancy PCA.Fancy PCA also improved Top-1 performance by 1.28%.
- The exact mechanisms producing different geometric-versus-photometric outcomes on coarse-grained data sets remain unresolved.The authors identify this mechanism as ongoing research.
6 Conclusion
The study concludes that generic data augmentation can improve CNN classification on coarse-grained data, especially through geometric transformations such as cropping. Transferability beyond Caltech101 and broader augmentation settings remain for future investigation.
- Geometric augmentation outperformed photometric augmentation on the coarse-grained Caltech101 data set.The evaluation used a relatively simple CNN architecture.
- Cropping generated the greatest reported gain, increasing Top-1 classification accuracy by 13.82%.The conclusion attributes the broader importance to transformations that alter image geometry rather than only lighting and color.
- Future work will test whether the Caltech101 results transfer to other coarse-grained data sets.The authors also plan to examine different CNN architectures, augmentation methods, and method combinations.