MIT installed AI on the laser engraving machine to automatically identify materials to determine the intensity of engraving, with an accuracy rate of 98%

rich color from the concave temple
qubit report | public account QbitAI

MIT recently published an interesting study:

span4_spanspan 4 laser engraving machine on the AI hybrid material Flowers can be carved on the T-shirt, and the power is automatically changed to ensure that it will not be cut

. They put an AI strong 24strong laser engraving machine on it, and it can automatically span 4 _span29 span4 _strong. Recognize 30 kinds of different cutting materials, accuracy rate is as high as 98% .

not only tells you what it is, but also tells you The force and speed required for cutting/engraving .

Of course, if this is a dangerous material, just throw a big "Caution" to you.

In this way, it can avoid some people from identifying mistakes /strong span_strong /strong 24strong /strong The material) , such as toxic smoke or something, and the cutting force is not correct to destroy the material waste caused by .

• is polycarbonate on the far right,Hazardous materials, do not (cutting will produce highly toxic flames) ;
• strong 24 laser can not be cut, but _7span strong laser can not be used for cutting 24 Alcohol wipe ;
• The leftmost is a transparent acetate sheet, which can be washed with warm water and mild soap, or with alcohol.

is a protective mask used in the laboratory. Of course, choose the leftmost one. Take it to get it done:

img3span

6 kinds of materials to be engraved at a time,It can match the design pattern with the corresponding material. After entering the thickness of each material, SensiCut immediately tells you that the felt is too thin, and the current pattern design is too complicated for it.

Then you can increase it a little bit before carving.

finished:

if you do not adjust as it says, to break out of it is "dragged down":

In addition, it can also perform laser-assisted engraving of patterns on mobile phone cases, clothes and other materials with mixed materials.

△

Note c red box portion, it automatically divides the design

this T-shirt by the following textile material and yellow sun part of plastic _strong29, material compositionThe middle of picture b shows the seagull pattern completed by SensiCut by identifying materials and then cutting strength to guide the completion. The effect is the best.

The hardware part is composed of laser pointer, lensless image sensor cutting head and battery, which are fixed in the 7span strong24span laser and battery. .

recognition principle is simple:

use speckle sensor (speckle sensing) technology, the laser To the surface of the material, the small feature difference above causes a small deviation of the reflected laser beam's optical path, which is reflected to the image sensor as a speckle pattern with bright and dark spots.

The following picture shows three sets of photos of the four materials under normal camera, electron microscope and speckle sensor imaging. The contrast is obvious:

img_sp12 spansp4 With the image, you can use the trained neural network for type recognition.

To ensure accuracy, the researchers trained 38000 images with 30 different material types.

They used transfer learning and pre-trained ResNet-50 _strong29 on the ImageNet dataset, and the model was optimized by 64Adam with a batch size of 0.00.

also used data enhancement technology to generate additional images so that the model is better generalized spanspan_span4 .

Evaluation results and future directions

SensiCut 98.01% (SD=0.20) strong span4 .

Paper material has the lowest accuracy rate _4span (silicone and leather are also easy to mix.

They also did some experiments and found that compared with the 100% recognition accuracy of red and white materials, the black material with less reflected light has only 92% accuracy, but this is when capturing images. Enable adaptive exposure to adjust.

In addition, they also studied the influence of light and angle on material recognition.

found that:

• increased brightness has no significant impact on the detection results of black/white materials, but transparent materials have a significant impact.After re-capturing images of transparent materials under different lighting conditions for training, the accuracy rate is 22% faster than the original.
• The angle that affects the most is wood, and the average detection accuracy of materials under 45% tilt is only 70.31%. This is because the cell 3D microstructure of natural wood has 90° rotational symmetry at the microscopic level.

Finally, the researchers said that in terms of hardware, all other components are available in the existing laser engraving machines, and the manufacturer _strong4span7span_strong4span7span_strong4span7span4 Lensless image sensors can have this technology.

In the future, the team will also evaluate how speckle sensing is used in to estimate the thickness of the material Further research in other aspects.

address of the paper:
https://groups.csail.mit.edu/hcie/files/research-projects/sensicut/2021-UIST-SensiCut-paper.pdf _spanspan4 reference spanspanspan_span7 Link:
[1]https://hcie.csail.mit.edu/research/sensicut/sensicut.html
[2]https://www.youtube.com/watch?v =1CjrVntolmo

— end —

qubit QbitAI · headline sign contract

follow us,Get the latest news of cutting-edge technology