0

I generated 700 1024x1024 images of female faces using an API. I labelled them either as attractive or unattractive.

The neural net should learn which face I find attractive and which not. But the accuracy is worse than a random guess.

To crop the background I used a cascade classifier from the opencv library.

def crop_background(img):
    np_img = np.array(img, dtype='uint8')
    face_cascade = cv2.CascadeClassifier();
    face_cascade.load(cv2.data.haarcascades + 'haarcascade_frontalface_default.xml');
    faces = face_cascade.detectMultiScale(np_img)
    if (len(faces) == 0):
        return img
    (x, y, w, h) = faces[0]
    face = np_img[x:x+w, y:y+h, :]
    return PILImage.create(face)

I use fastai to build a model quick.

dls = ImageDataLoaders.from_folder(
    path, get_image_files(path), valid_pct=0.2, seed=42, item_tfms=[Transform(crop_background), Resize(224)], bs=4)

I used a pretrained resnet34 as the architecture.

learn = vision_learner(dls, resnet34, metrics=error_rate)
learn.fine_tune(2)

I expected the accuracy to be around 90%. But the accuracy is around 50% not better than random guessing.

I thought about generating more data or running for more epochs or using a different architecture.

Erdogan Seref
  • 1
  • Please clarify your specific problem or provide additional details to highlight exactly what you need. As it's currently written, it's hard to tell exactly what you're asking. – Community Dec 04 '22 at 17:08
  • You may need to add more data for training and validation during training, use more layers and tune the parameters. Also use more metrics for scoring the model – Dhana D. Dec 06 '22 at 05:17

0 Answers0