CODEWITHSUNDEEP
pythontensorflowkerasneural-networkconv-neural-network
Deshwal
Deshwal

Reputation: 4152

minimum image dimension (w*h) too small in the dataset for it to use as target_size in Keras CNN

I have very robust data in my dataset varies in dimensions, intensity,flip, shear etc. By using, 

from PIL import Image
mini = 99999

for _,_,files in os.walk(train_dir):
    print(len(files))
for pic in files:
    with Image.open(train_dir+pic) as img:
        width, height = img.size
        if min(width,height)<mini:
            mini = min(height,width)
print(mini)

it prints 19996, 8 and it means that either the minimum of height or width of all the images is 8. So can I use

ImageDataGenerator().flow_from_dataframe(target_size=(8,8) ). I think it won't be a good idea to use this. and if not (8,8) what should I use? I am using the data for Convolution Neural Networks for Age Detection in Keras using Tensorflow as backend.

Upvotes: 0

Views: 565

Answers (1)

lenik
lenik

Reputation: 23538

Please, do this instead:

min_w, min_h = 10000, 10000
for pic in files:
    with Image.open(train_dir+pic) as img:
        width, height = img.size
        min_w = min( min_w, width)
        min_h = min( min_h, height)

print(min_w, min_h)

Then you will have at least reasonable understanding about the smallest image size.

Regarding the second part of your question -- NO, target size of the image of 8x8 is too small to do anything meaningful, so you may want to discard the images that are too small or upscale them or do something else, depending on the data contents.

I would consider the image size about 100x100 ~ 200x200 a good size for practicing with CNN in Keras.

Upvotes: 1

Related Questions