YOLO DNN trainning/ Entrainement d'un réseau de neurones YOLO

I am currently working on a robotic challenge in which we need to detect specific patterns. I found out that the easiest way to do so is to use a deep neural network, the one I use is YOLO(You Only Look Once). It can detect object on an image and their position.

Here it is a link to describe the whole procedure to train a YOLO DNN to recognise specific objects (in french):
https://drive.google.com/file/d/1S6Xmu4muZuLkwgQUTueVIePNKMEK9qIs/view?usp=sharing


Je travaille actuellement sur un défi robotique dans lequel nous devons détecter des objets avec une camera. J'ai trouvé que le moyen le plus simple pour le faire est d'utiliser un réseau de neurones, celui que j'utilise, c'est YOLO (You Only Look Once). Il peut détecter des objets sur une image et leur position.

Voici un lien pour décrire la procédure complète pour entrainer un réseau YOLO à reconnaître des objets spécifiques:
https://drive.google.com/file/d/1S6Xmu4muZuLkwgQUTueVIePNKMEK9qIs/view?usp=sharing

Résultat de l'entrainement

LCD polarization

LCD polarization

I recently found a old LCD display and decided to use the glass layer for future experience. But when I disassemble the screen the polarization of the screen was so beautiful, here the picture:

 On the last picture you can see the ITO (indium-tin oxide) conductive layer, which correspond to the inscription on the screen. I was surprise to measure a conductivity of only 50 Ohm per centimeter, this measure is inaccurate but I though it was more conductive.

CO2 Laser: What changes since the beginning?

CO2 Laser: What changes since the beginning?

Since the beginning of the project, this winter, a few things change.

Here a quick summary:

    The first version was maybe the simple one, few copper tubing, a glass tube and two alignment plate. Initially the two connectors were meant to pump down the system and and the gas mixture into the tube. But this design is silly because the mixture added is directly suck up by the pump and didn't go on the whole tube. Actually to fill the tube, you need first to stop the pump, fill the tube and then pump down again the system. The easiest system found to hold the tube are three curtain holder, they are isolating and strong enough to hold it.

Note: On this drawing the glass tube appear quite short but it is to facilitate the representation. On the real tube the tube length is about 80cm. 

   The second version is like the first one, except for the water-cooling system and the support system. The water cooling system is just an acrylic tube with two small tube connect to it. I use a small aquarium pump to circulate the water. The support is two pieces of wood, stronger than the curtain holder, because the curtain holder wasn't rigid enough, precise adjustment was therefor impossible. The connection between the copper tubing and the glass tube also evolved, before the tightness was insure by tape, but there were tiny leak between each tape layer. So it was replaced by silicon join, witch extremely reliable and tight. Unfortunately, the plasma start exactly on the junction between copper and glass near the silicon and burn the silicon. I think silicone vapor are at the origin of the plasma conductivity and color changes. Moreover when I start the plasma the pressure rise dramatically.

   The third version should solve the previous issue, but pushing away the silicon the plasma start line, as you can see it on the image. The mirror alignment system has been installed. Unfortunately, the plasma cover the whole copper part and continue to burn the silicon.

CO2 Laser: Custom alignement system

CO2 Laser: Custom alignment system

Mirror alignment is crucial for a laser. To align one the mirrors you can simply look thought the tube and screw or unscrew the three blots on the alignment plate to see the reflect of your eye in the mirror. But to align the second mirror you need to remove the end plate with the align mirror.

     So the alignment system allow the align of the both mirror from outside the tube once you align each one from inside.

On the image above, the third version of the laser you can see on the end plates one laser (gold) and one spike.

Misaligned mirror

When you align the center of your eye with the spike and the red dot on the mirror, you can directly see how align is the mirror. On the example above the mirror is misaligned because the reflect of your eye (the black dot), the reflect of the spike and the red dot are not align.

Aligned mirror

Once they are aligned, as above, you can see that the three part are merged.

     I think we can say that system is quite precise because between the two images the mirror angle is only of 1/2 degree, the real system is even more accurate because the glass tube is much short than on the real system.

Vacuum chamber part three

Vacuum chamber part three

I had some hard time with feed-thought connection, so I decide to solve this issue. To do so I use epoxy resin. I wanted a lot of electrical connection for connecting all the instruments, the high voltage connector for sputtering and high power connector for metal evaporation. I use the following template for the different connectors.

The connector

The mold with resin pour inside

The result

On the plate

As you have maybe notice, there is no widows on the chamber so I cannot see inside except if I put a webcam inside to get in real time what going on inside.

The connector with USB connection

On the left you can see something which look like a light-bulb, this is my pirani gauge, unfortunately the wire to the measuring circuit is too long so the measure aren't usable.

The setup

 On the picture above I tried to boil and freeze water. It only boil, I haven't waited long enough for it freezes.
 

Vacuum chamber part two

Vacuum chamber part two

Now that the base plate and the chamber are done I need to make the connection between them vacuum-proof. To do it I choose one of my favorite material the acetic silicone join, it is easy to work with, totally vacuum-proof. The hard part of join making is to make an even join but no that thick. If it is the pressure the chamber edge on join will cut it. The total pressure exerted on the chamber edge is about 10 tons, quite huge! :) .

The cardboard is so useful 

Finish join

The join you can see above have been cut apart a few time later and replace by a thinner one.

Once the join was finish I've solder the connection port to the vacuum chamber.

Vacuum chamber part one

Vacuum chamber

     After playing a bit with vacuum technologies and plasma I've found several problem on my setup. The base plate deform no is no longer flat after soldering. Also electrical connection which pass through the plat leak. Furthermore, the bell jar is in glass and don't really trust it, I think it can implode. A previous jar crack went I lower the pressure inside hopefully it didn't implode. But meanwhile I found this site. So I decide to reproduce his setup with some modification to solve sealing, imploding and leveling issue.

       I first needed an empty gas bottle and I only had a half-full one. Given that propane or methane are flammable and also greenhouse gases, I've decide to burn it for security reason, I don't really want my neighborhood to explode. According to wiki, methane greenhouse effect is 70 time more important than CO2 one.

      So I decide to make a gardening flame thrower with only a gardening hose and a copper tube. I don't use any security valve or backfire one. I thought that backfire need oxygen contamination to occur, so I simply use the bottle valve to give a sufficient flow of gas. I use it to kill hen lice a plague for hen-house.

Burning lice

Once I empty the bottle I've decided to cut it a quick making is enough.

The bottle use with the mark line

After cutting in two part, as planned I found a dark smelly crust. I try to get rid of this crust by sanding it with sandpaper by it wasn't enough so wash it with high pressure water.

The gas smelling dark crust

I've try to level the bottom of the chamber with the grinder by i was inaccurate for small leveling.

     For the bottom plate I use a 5mm thick plate screw on pinewood plank, like this I hope to reduce heat deformation of the metal plate.