PiSegment

A lightweight Python software for semi-supervised segmentation on images, it can be utilized for tasks such as image segmentation, background extraction, semantic segmentation, colorization, etc. The code, written in pure Python, can be easily integrated with Flask or Django backend for a web app.

Results

Image w annotation	Segmentation

Segment 1	Segment 2

Video

out-web.mp4

Installation

pip install pisegment

Usage

• Command line

pisegment --input "path/to/image/tobe/segmented" --mask "path/to/the/generated/annotation"

• The above command opens up a GUI for image annotation. Use right click drag for annotation. There are 9 different color options. Press 1 for black color, 2 for blue color and so on. After annotating, press s to let the segmentation begin. Zoom in/out with mouse wheel and left click drag for panning image.

Tips on using:

• By default, the denoising filter is on. If your image doesn't have noise, turn it off using the --no_filter option. You may also use this option if you are already using different software for denoising.

pisegment --input "path/.." --mask "path/.." --no_filter

• For fast processing, consider downsizing your image to under 256 X 256 to generate the segmented mask, then upscale the segmented mask to the original size.
• For complex images, like the last example in the Results, to get a precise segmentation, the parameter --sig plays a crucial role. See the Jupyter_demo for the workflow on how to segment such kind of images.

How it works?

Behind the scenes, it is basically using a Dijkstra like algorithm to propagate the labels on graphs. The code is very similar to the Algorithm 2 of the following paper: https://hal.science/hal-00932510/document

Params Description

The following description of parameters is useful for achieving good segmentation.

Param	Description
sig	An important parameter for achieving good segmentation on complex images (see Jupyter_demo). By default, I am using a naive estimation of this parameter which may need to be fine-tuned.
no_filter	Turns off the denoising.
ps	Patch size, used for creating a non-local kNN graph for denoising. The default value is 3, which should be fine.
k_	The $k$ value in the kNN graph for denoising. The default is 10.
k	The image grid graph $k$ value. 4 or 8 should be sufficient. The default is 4.

pisegment --input "path/.." --mask "path/.." --sig 1.00e-02  --k 4 --k_ 10 --ps 3 --no_filter