[tensorflow & pytorch] image data enhancement API

Preface ：

In deep learning training , The training effect is poor 、 Small number of training sets 、 When there is a fitting trend, you can choose to increase the number of data sets to optimize the training model , But most of the time , The amount of time and effort spent on increasing the number of data sets is enormous , Therefore, the more common method we use is to enhance the existing data sets . It's better to actually increase the number of data sets , But it still has certain effect , High cost performance .（ Just add a few lines of code ）

TensorFlow Of API stay image Next ：（ I am using 2.0 edition , Different versions may API Different , But basically, it can be in iamge Find below ）

Module: tf.image  |  TensorFlow Core v2.0.0https://tensorflow.google.cn/versions/r2.0/api_docs/python/tf/image

from tensorflow import image

PyTorch Of API stay torchvision Of transforms Next ：torchvision — Torchvision 0.13 documentationhttps://pytorch.org/vision/stable/index.html

from torchvision import transforms

Here are some methods I think are commonly used . 

To be detailed API Click the link above to view the official documents .

Randomly flip the picture ：

TensorFlow：

Randomly flip up and down ：

output=image.random_flip_up_down(input)

Randomly flip left and right :

output=image.random_flip_left_right(input)

 PyTorch:

Randomly flip up and down :

layer=transform.RandomHorizontalFlip( )

Randomly flip left and right :

layer=transform.RandomVerticalFlip( )

Rotate the image randomly |n°|:

layer=transform.RandomRotation(n)

Notice that what you get here is a layer (layer), You also need to stuff the picture into the layer to get the result . and TensorFlow The example in is that the result can be obtained directly .

example :output=layer(input)

Crop the image randomly :

TensorFlow:

take input Cut randomly into shape size :

output=image.random_crop(input,[shape])

take input Random cutting , The central part is reserved central_fraction(0~1). for example central_fraction=0.5, Then half of the central part is reserved .

output=image.central_crop(input, central_fraction)

PyTorch:

Get to crop the incoming image into shape The layer :

layer=transforms.RandomCrop(shape)

Get the central part of the incoming image and crop it into shape The layer :

layer=transform.CenterCrop(shape)

Regularization :

Strictly speaking, regularization is not data enhancement , But it can make the image more convenient for training . The following example is limited to three channels RGB Images .

Change the pixel value of the image into an average of mean, The variance of std Is a normal distribution .

Following mean and std It is the optimal value calculated by predecessors ( Maybe ?)

TensorFlow:

TensorFlow There seems to be no interface , So I have to write it myself :

mean=tf.constant([0.485,0,456,0.406])
std=tf.constant([0.229,0.224,0.225])
output=(input-mean)/std

PyTorch:

 layer=transform.Normalize(mean=[0.485,0.456,0.406],std=[0.229,0.224,0.225])

ending : 

TensorFlow Of image There are many image processing API( Not limited to data enhancement ), If you are interested, you can have a look at .( I'm tired of reading English if I'm not good at it )

Because of the above PyTorch What you get is layers , So it can be packaged into a container :

transform.Compose([
            layer1,
            layer2,
            layer3,
            ……………… 
            transform.ToTensor()# Finally, remember to add this layer , Turn the image into tensorf Can be pytorch distinguish 
        ])