[PYTHON] Face image dataset sorting using machine learning model (# 3)

Overview

Sort UTKFace datasets by the following features. ·no problem ·grayscale ・ Multiple people are shown ・ Processed

Sorting flow

Use Last created learning model to identify the image data in the part1 folder of the UTK Face dataset one by one. ↓ Create the following folder in Google My Drive and save the identified image ・ Normal ・ Grayscale ・ Several ・ Processing

environment

Google Colaboratory（GPU）

#UTKFace part1 ZIP read
z = zipfile.ZipFile('/content/drive/My Drive/part1.zip')
imgfiles = [ x for x in z.namelist() if re.search(r"^part1.*jpg$", x)]

for imgfile in imgfiles:
    image = Image.open(io.BytesIO(z.read(imgfile)))
    image = image.convert('RGB')
    image_p1 = image.resize((im_size, im_size))
    data = np.asarray(image_p1)
    #shape(299,299,3)→(1,299,299,3)Change to
    data = data[None, ...]
    #Get the file name saved in the ZIP folder
    file = str(os.path.basename(imgfile))
    #Predict by inputting data for one image into the trained model
    pred=model.predict(data)
    #Change the location to save the image data according to the determined value
    if np.argmax(pred) == 0:
      image.save('/content/drive/My Drive/normal/' + file)
    elif np.argmax(pred) == 1:
      image.save('/content/drive/My Drive/grayscale/' + file)
    elif np.argmax(pred) == 2:
      image.save('/content/drive/My Drive/several/' + file)
    else :
      image.save('/content/drive/My Drive/processing/' + file)

z.close()
del z, imgfiles

Since model.predict assumes reading of multiple images (model arguments are 4D), an error occurred when trying to predict one by one. He predicted that data = data [None, ...] would convert from 3D to 4D. The image file name is not changed and is sorted and saved in each folder.

Check the images in the sorted folders

(I put a screenshot on it, but I deleted it because I thought it might be illegal.)

grayscale folder

You can see some colored images here and there, but it seems that you can almost distinguish them.

several folders

I was learning with images showing multiple people, but probably because there was little learning data, I feel that images with stuffed animals and animals, or images with a non-simple background are also in this folder. I did.

processing folder

It is possible to almost distinguish data that has characters on the face or a swirl-like pattern.

result

Although the discrimination accuracy is uneven, I was able to collect neat images in the normal folder. I will use this normal folder to create a model that can identify age and gender.