Is there no way to make jitterColorHSV work within augmentedI​mageDatast​ores?

2 次查看（过去 30 天）
Dominique Chabot
Dominique Chabot 2020-10-3
编辑： Dominique Chabot 2020-10-3
I find the imageDataAugmenter function and the derived augmentedImageDatastores very handy for training deep neural networks, as they can generate a virtually infinite number of random variations of the training images that are changed at each epoch during training, without permanently modifying the original underlying training images.
I only wish that in addition to the reflection, rotation, scaling, shearing and translation modifications that are currently supported by imageDataAugmenter, you could also get it to perform the jitterColorHSV operations.
Because as far as I can tell, if I want to "augment" my training images by introducing random color jittering, the only way to do it is to permanently modify them using the jitterColorHSV function, which will yield a single set of variations that will be reused at every training epoch, which isn't nearly as powerful as what augmentedImageDatastores do with the imageDataAugmenter.
Or perhaps at best I could make several duplicates of my training image set, apply random color jitters to each copy (e.g. via transform), and then combine them together in a single imageDatastore for training, but then I'd be creating new issues by oversampling the training data: i.e. the model would see several variations of each training image within a single epoch, and furthermore it would see the same set of variations at every epoch; instead of one new variation of each training image at every epoch.
On the off-chance that someone knows some sort of clever workaround to get jitterColorHSV to work similarly to augmentedImageDatastores, I would greatly appreciate hearing about it. Otherwise, if any MathWorks staff happen to be reading, please consider this a feature improvement request.

请先登录，再回答此问题。

回答（0 个）

类别

在 Help Center 和 File Exchange 中查找有关 Image Data Workflows 的更多信息

标签

产品


版本

R2020b

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!

Translated by