数据增广（data augmentation）在训练深度神经网络时的必要性解释

Left-Ventricle Quantification Using Residual U-Net

Kerfoot E., Clough J., Oksuz I., Lee J., King A.P., Schnabel J.A. (2019) Left-Ventricle Quantification Using Residual U-Net. In: Pop M. et al. (eds) Statistical Atlases and Computational Models of the Heart. Atrial Segmentation and LV Quantification Challenges. STACOM 2018. Lecture Notes in Computer Science, vol 11395. Springer, Cham. https://doi.org/10.1007/978-3-030-12029-0_40

3.1 Image Preprocessing

An accurate and robust segmentation network requires a large dataset so that it learns a general solution which can correctly segment images not seen in training, and does not become over-fitted to the input data. The key concept is variety since the network is being trained to identify geometry embedded in varying contexts. If some ancillary feature in this context is often present with important features the network will correlate these features and produce a poor result for inputs lacking the ancillary feature. To provide this varied data, we first create an expanded input dataset from the original challenge data and then apply data augmentation during the training process.
…

During training, we use data augmentation^{1 2} when creating an input batch of images at each step. Images from the expanded dataset are selected at random, then a random selection of transpose, flip, 90/180/270-degree rotation, and shift operations are applied to each image/segmentation pair. Applying these transformations essentially produces a further expanded dataset which contains increased image variation although not further geometric variation. This prevents the network from fixating on features in specific regions of its perceptive field since the random transformations move such features around the field during training. Augmentation is the only technique we use to prevent over-fitting, other techniques like dropout were found not to improve performance and so omitting them contributed to a simpler network architecture.

---

U-Net: Convolutional Networks for BiomedicalImage Segmentation

3.1 Data Augmentation

Data augmentation is essential to teach the network the desired invariance and robustness properties, when only few training samples are available. In case of microscopical images we primarily need shift and rotation invariance as well as robustness to deformations and gray value variations. Especially random elastic deformations of the training samples seem to be the key concept to train a segmentation network with very few annotated images. We generate smooth deformations using random displacement vectors on a coarse 3 by 3 grid. The displacements are sampled from a Gaussian distribution with 10 pixels standard deviation. Per-pixel displacements are then computed using bicubic interpolation. Drop-out layers at the end of the contracting path perform further implicit data augmentation.

---

1. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. In: Proceedings of the 25th International Conference on Neural Information Processing Systems, NIPS 2012, vol. 1, pp. 1097–1105. Curran Associates Inc., USA (2012) ↩︎

2. Simard, P.Y., Steinkraus, D., Platt, J.: Best practices for convolutional neural networks applied to visual document analysis. Institute of Electrical and Electronics Engineers, Inc. August 2003 ↩︎