hello Guys.
Is it possible to use Deep Learning or Computer Vision to detect a Weld present in the Image. I would like to location a Weld in the Image, how can i do this? I have attached an example here,
1 Like
I am a noob as well, so I couldn’t tell you much about the how.
But from what I have gathered from Lesson 1 and 2, if a human can accurately distinguish an image with a weld, then a Deep Learning model should be able to do so as well.
Fairly confident, that using the code from lesson 1, you should be able to train a basic model that can identify if an image contains a weld. How to take it beyond that to labelling the part of the image with the weld is currently beyond me.
Still, there are Youtube videos on how one can get AI to detect hands from a webcam feed and then label the image in real time so that the fingers are rigged. You might be able to use a similar mechanism to label your image.
1 Like
hi there!
You have several options to do that.
The first is using object detection, if it is really just about “is there a weld?”
For this, the labeling consists out of having a few (hundred) images with welds, drawing boxes around them, letting a network (yolo for example) see enough of them - and Bob is your uncle.
Or you go the segmentation way, the labeling gets more complicated since you are not doing boxes, but the exact shape of the object, often times pixel by pixel. The advantage with this, provided that you know the distance of the camera to the part, is that you can calculate the width and length of your weld and more likely detect issues.
This video is one of the most recent ones using fastai for image segmentation
But you can also look that up by Jeremy himself, in the 2019 lecture series I think he did this topic as well.
1 Like
i see , thanks a lot!!
1 Like
ok thanks for sharing the Info. I will get to work…
Np fam, glad to help. :3
(Segment Anything) SAM Model can help you build your segmentation dataset,
2 Likes
if i train a Model based on Images, would it be capable of detection from Video stream?
or is it possible to train a Model based on Video stream using SAM Model or U-Net?
Depends, Does it need to be real-time or will you just process a video file frame by frame? for real-time you’d need the model inference to be quick enough to keep up with the framerate, on the other hand if just using the video as a collection of images (frames) then you can use tools like opencv to iterate through the frames, extract the image and present to a slower model.
Supervision looks to have a lot training and tools in-place for object detection, and video.
Check out some of their videos at the bottom of the page. Might be worth some exploration there. (I’ve not used it beyond a trying a few demos, but hope play more when I have time)
1 Like