Keras: Converting a JPEG to a form in which it can be send into a neural net

[AntreasAntoniou]

Hello there,

I am experimenting with keras and imagenet , my network is setup so is everything else, I am just making sure that my inputs are correct by asking you here. So basically I take the JPEG images and resize them , change their format from BGR to RGB and finally using numpy change their shape so they would fit the shape of the network (samples,3,224,224), when I load them with OpenCV the format is 224,224,3 so I have to reshape them , but I am afraid that numpy may not be reshaping them correctly. Can you tell me if what I am doing looks correct?

file=cv2.imread(filename)
file=cv2.resize(file,(224,224))
file=cv2.cvtColor(file, cv2.COLOR_BGR2RGB)
file=np.array(image).reshape((3,224,224))
files.append([file])
print(file.shape[0])

[AntreasAntoniou]

Hey , for anyone having the same question I managed to fix it. There was indeed an issue with the way numpy reshaped things, I found a better way to do it. Simply extract all 3 channels separately and put them in a new array which changes the shape very simply. Here is a piece of code that showcases the 2 different approaches , the numpy reshape and the manual reshape. The first one gives a weird output whereas the second produces the correct image.

import numpy as np
import random
import cv2
import matplotlib.pyplot as plot
from skimage import data

image=data.hubble_deep_field()

image=data.astronaut()
image=cv2.resize(image,(224,224))

image=cv2.cvtColor(image, cv2.COLOR_BGR2RGB)

array=np.array(image)
print(array.shape)
print(array)
new_image=np.array(image).reshape((3,224,224))
print(new_image.shape)
print(new_image)
plot.imshow(image[:,:,1])
plot.show()
plot.imshow(new_image[1,:,:])
plot.show()

red=image[:,:,0]
green=image[:,:,1]
blue=image[:,:,2]

image_convert=np.array([red,green,blue])
print(image_convert.shape)

print(image_convert)

plot.imshow(image_convert[1,:,:])
plot.show()

[arogers1]

You can use Numpy's swapaxes function to reshape the image into Keras' format in one line of code.
http://docs.scipy.org/doc/numpy-1.10.1/reference/generated/numpy.swapaxes.html

import numpy as np
import random
import cv2
import matplotlib.pyplot as plot
from skimage import data

#read image into cv2 format
image=data.astronaut()
image=cv2.resize(image,(224,224))

#use swapaxes to convert image to Keras' format
image_convert=np.swapaxes(np.swapaxes(image, 1, 2), 0, 1)

#Show that it worked
print(image_convert.shape)
plot.imshow(image_convert[1,:,:])
plot.show()

[sharoseali]

can u please tell me ..what is (image, 1, 2), 0, 1) in the code...?

[SomyaNIGAM]

Hi!
There is one issue that I am facing, I need to swap the dimensions of the image for the neural network. But, doing that changes the colour of the images. Is there any way to get this done?