Hi Yacine, I had to use a lot of pap cell images to train my network and three methods work well: 1.- Use paint to change size and position. 2.- Use power point to rotate the images one degree at a time. 3.- Use openCV to add white noise to each image. Good luck.  

Following can work only if your image data set is encoded/encrypted and saved with appropriate pixel info. 

You can use data pre-processing options available in WEKA software.

Open your data file in .ARFF form. Click the preprocess tab. Choose 'filters'.

Under this you select 'supervised' if you have the output classes else select

'unsupervised'. Select 'attribute' under this you have several techniques like

'replace missing values' OR 'replace missing values with user constant' or 'add values'.

This is to fill-up a few less important missing values. In case, if you wish to increase the

whole data size you may use my bootstrap code written in MATLAB.

data1 = ([96, 97, 98, 99, 102, 110, 130, 140, 170, 178, 182, 210, 222, 225, 230, 240, 245, 260, 270, 271, 273]);

[n n1]=size(data1);

a=mean(data1)

b=std(data1)

m = bootstrp(5000,@mean,data1);

d=mean(m);

e=std(m);

In the above code, data1 is initialized to the actual data available.

Value 5000 is the number of data samples you need. Note that this code

is only for generating multiple values for a single attribute/parameter.

For example, you have about 21 actual values for an air pollutant PM10

for a given location. The above code will help you to generate more values about

the mean. Run the code in MATLAB and type 'm' on the prompt to see the generated values.

'a' to see the mean and 'b' for standard deviation. You may decrypt the image data set.

I think you could try some bayesian networks in some small size datasets. If your dataset is large-scale, CNN has a good performance in terms of accuracy. But if the dataset is small, CNN tend to be overfitting.  I suggest you to use some 'older' methods like SVM, random forest,etc. 

Hi Yacine, I had to use a lot of pap cell images to train my network and three methods work well: 1.- Use paint to change size and position. 2.- Use power point to rotate the images one degree at a time. 3.- Use openCV to add white noise to each image. Good luck.  

Hi,

I would like to thank you for all your answers, they will help a lot.

@Kalyani Sambhoo: I perform the image processing with Python (OpenCV+Dlib) but the idea generating new images from statistic measures is interesting : mean and std images (with a certain size of window)

@Pengbo Zhang: I already used other classic methods to do that and I got pretty interesting results. Actually, I just took a challenge of exploiting the CNN in order to do facial expressions recognition and there are some issues to wich I need to find solutions.

@Craig Michael Thomas: Using mirror image is a very good idea, I forgot that I read in a paper. With your suggestion, if I use horizontally and vertically i can increase the size of the dataset x4 times. Thank you.

@Oscar. Chang: I can't change the size of the input image because I defined my network based on a certain size of the image. The position is not an add value step because the convolutional layer process and take into account the position. But the rotation can be an interesting idea for generating others images.

Thank you very much for all your answers.

Yacine.

Hi, Yacine Yaddaden,

If you would like to add some additional data, it may be a good idea. But you must make sure that the distribution of additional data is identical to the original data. 

Besides, in my opinion, you can use some regularization methods like dropout to avoid over-fitting. But I think it is not an easy thing.

@Pengbo Zhang: The distribution will be the same if I apply the same transformation for every single image of the training set. I've read some paper about that, the concept is very simple, it just ignore some nodes of the hidden layer where it is applied with a certain probability (usually 0.5). I'm using Torch7 as Deep Learning framework, fortunately for me, it's included as a function. I'm testing it.

Thank you for the suggestion of using such a function.

Yacine.

Hi, Yacine Yaddaden

I am very glad to hear that my advice may help you, and you also can try other methods to assist you. 

A lot of good answers!

I have been using data augmentation with reasonable success for small dataset working with high-capacity CNN. Some frameworks support online image augmentation while running the training. For example, in Caffe, mirroring and random cropping are possible without changing actual DB.

Other pre-processing based augmentations that are non-destructive of ground truth (don't required changing ground truth labels): change saturation, change contrast, blur/sharpen image -- especially to mimic camera shakes, etc.

Augmentation that may require ground truth updates (e.g. if labels contain bounding boxes): rotation, perspective distortion, panning. These are tricky as ground truth also needs to be updated procedurally to match the operation on the image. However, these do produce robust increase in dataset.

Hope this helps!

You can try Image flipping,random cropping,random scaling,centre zooming,increasing/decreasing brightness,sharpness, introducing slight noises into the image pixel values.

maybe late but image augmentation would be easier which includes basic transformations and color enhancements.

I am using MATLAB for object detection and i have small dataset with greylevels images i want using data augmentation to avoid overfitting but i dont know how can i add this in the algorithm because the input of object detector are the images and the groundTruth with bounding boxes so how can we add data augmentation to labeling bounding boxes ????

Resamplig the data set

Thank you for your answer. Howner can i do it in m'y case i want data augmentation of labeling ground Truth ??

can I use flipping,rotation and other augmentation methods for a dataset having medical images? @oscar.chang@pengbo zang

All the answers are great

Keras Provides builtin tool for augmenting data, the data generator, it has number of operations which can be use to change the image mathematically with same information like rotation, flipping, zooming, blurring, scaling and lot more.

Data augmentation is the best to increase the feature for the training of the learning models. For data augmentation rotate the image at a different angle which increases the size of the dataset.

There is new wave toward automated data augmentation to improve the generalization performance. For instance, this paper https://arxiv.org/abs/1906.11172 utilizes reinforcement learning to create an opimal data augmentation policy for object detection.

You can use Data Augmentation:

https://imgaug.readthedocs.io/en/latest/

Besides Data augmentation, I would suggest using same domain transfer learning. It is better than data augmentation techniques in some applications

For more details, check this paper

https://www.mdpi.com/2076-3417/10/13/4523

Hello, eveyone, if i use the data augmentation in keras, but the loss of model when train it still so big, what your suggestions? # if i have imbalance dataset