Alpha blend

We will attempt to perform semantic segmentation of the original image. The Semantic segmentation will be alpha blended with the original image for SLIC segmentation.

Semantic Segmentation

Implementation 1: https://towardsdatascience.com/image-segmentation-with-six-lines-0f-code-acb870a462e8

1. Install pixellib: pip install pixellib Had to go through a re-install of Python. Now running on python version 3.8.

Have the sample segmentation code running.

The baseline version of the code runs against the model deeplabv3_xception_tf_dim_ordering_tf_kernels.h5. This can be downloaded here: https://github.com/ayoolaolafenwa/PixelLib/releases/download/1.1/deeplabv3_xception_tf_dim_ordering_tf_kernels.h5 This seems to have a very limited number of classes.

deeplabv3_xception65_ad320k.h5 performance

Seems to perform better: https://morioh.com/p/b04406d57772

Model download: https://github.com/ayoolaolafenwa/PixelLib/releases/download/1.3/deeplabv3_xception65_ade20k.h5

This semantic segmentation method performs best when it finds an object from its training categories, e.g. a horse. The segmentation create far more segments than expected. However, if it is an object it does not recognize, it does some odd region identification. Although not completely correct, it may still prove useful.

mask_rcnn_coco.h5 performance

Additional tutorial: https://www.analyticsvidhya.com/blog/2019/07/computer-vision-implementing-mask-r-cnn-image-segmentation/?utm_source=blog&utm_medium=introduction-image-segmentation-techniques-python

Model download: https://www.google.com/url?sa=t&rct=j&q=&esrc=s&source=web&cd=&ved=2ahUKEwjSntaMnNjrAhVIl3IEHe_0DV0QFjAAegQIAhAB&url=https%3A%2F%2Fgithub.com%2Fmatterport%2FMask_RCNN%2Freleases%2Fdownload%2Fv2.0%2Fmask_rcnn_coco.h5&usg=AOvVaw0nAUAmHpcXDQ6mPgV9NckR

This model only highlights the objects that it recognizes.

Reducing noise of semantic segmentation

The deeplabv3 semantic segmentation can produce many smaller segments. For example, this segmentation image.

Test reducing the number of regions by testing these filters:

1. Blur filter

Blur 10x10

Blur 20x20

2. Bilateral filter

Bilateral(9, 75, 75)

Bilateral(20, 75, 75)

Bilateral(50, 75, 75)

3. Median filter

Distance 9

Distance 19

Distance 31 --> min( ) / 20.

Distance 61 --> min( ) / 10.

Distance 121 --> min( ) / 5.

Median filter with min ( ) / 10 seems to work reasonably with deeplabv3.

Pebble mosaic

https://link.springer.com/article/10.1007/s41095-019-0129-0