What are possible approaches for recognizing rotated numbers as well with TF API?

General Discussion
#1

I have read some examples of NNs for image recognition, however, I can’t find much about making this recognition rotation invariant.

Is there any API that could be of help in doing it on the fly ? I am wondering also how difficult would it be when the labels change, like when doing bounding box.

#3

@Mah_Neh Welcome to Tensorflow Forum!

Here are approaches you might want to try for your usecase to achieve rotation invariance in image recognition, along with considerations for bounding boxes:

During training, create multiple rotated versions of each image (e.g., 90, 180, 270 degrees). Update bounding box coordinates accordingly for rotated images.

Operate on spherical data representations for inherent rotation invariance. Capture spatial relationships and pose information more effectively than traditional CNNs.

Learn to spatially transform input images, including rotations, for better alignment.

Update bounding box coordinates based on the rotation angle applied to the image.

APIs for Rotation: -

  • OpenCV: cv2.rotate()
  • Pillow: Image.rotate()
  • TensorFlow:tf.keras.layers.RandomRotation() , tf.image.rot90()
  • PyTorch: torchvision.transforms.RandomRotation()
    fort.

Let us know if this helps!