Bordering images with python Part 1

What you want to do

There are many opportunities to outline the creation of machine learning datasets, but I want to automate it as much as possible.
Learning openCV, numpy, matplotlib
~~ The Lord, who has a strong Miriota component, knew how to induce torpedoes (Wake Homing), so I wanted to use it. ~~ As I played around with it, I lost the wake homing relationship.

Well, I will summarize what I tried variously like this as a memorandum.

environment

ubuntu 18.04 LTS
Anaconda 1.7.2
python 3.7.3
openCV 3.4.2
numpy 1.16.2
matplotlib 3.03
Jupyter Notebook

Prerequisite knowledge

How to operate python and each library
Polar math

Image for testing

For the time being, I thought that it would be possible to do RGB images if I bite the image processing, so first I will use binary images. I challenged to take the edge of a complicated figure because the inside is missing.

code

Github Jupyter Notebook

Basic policy

Make a dot inside the object you want to demarcate
Draw a line outward from that point and find out where on that line the color changes.
Rotate the line 360 degrees to get the boundary.

A simple image looks like this

↓ Detection target Detection result test01検出結果01.png It's almost taken.

However, depending on the position of the point, it becomes a shadow of the line and looks like this. test01検出結果02.png

I would like you to understand that the image at the beginning is a rather difficult task.

Test image detection result

test検出結果01.png test検出結果02.png test検出結果03.png test検出結果04.png test検出結果05.png

With one point, shadows can be created everywhere, and invisible areas can be created.

If there are multiple points, the shadow problem should be alleviated.

test検出結果06.png

Overdone w

Summary

The outer edging is roughly removed. Still, there are some areas that cannot be removed in the internal area, so improvement is required. But I don't think it's difficult to border with a Pezier curve with this much.

This time, I hit the dots randomly, so there were too many wasted hits, but it seems interesting to move the dots appropriately and remove the edges with the minimum number of dots. I wonder if I can use reinforcement learning. It will be an issue from the next time onwards.