How do I apply Object Detection with Deep Learning to a relatively simple problem?

2 次查看（过去 30 天）

Jannis Holtkoetter 2020-7-18

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/567111-how-do-i-apply-object-detection-with-deep-learning-to-a-relatively-simple-problem

评论： Jannis Holtkoetter 2020-7-18

Hello everybody, I am relatively new to MATLAB and would like to solve following problem:

I want to create a script, that detects how many pills are left over in a "blister" (I am referring to the pill packages, where one has to pop out the pills). The images of the blisters are images which are taken with a smartphone and have random backgrounds (most of the times a table or a hand) and vary in the lighting conditions or angle. Moreover, the blisters are from different manufacturers and have different colors, different number of pills, some are transparent, some are very light-reflective. I do not want to detect these features with my program, I only need the number of pills. However, these are the factors which make the script more complicated.

So far, I am trying my best with a script which includes several image processing functions. The script can seperate the background from the blister in most of the times (~90% of the times). I can also detect and count the pills in the blister. However, the detection and counting of the pills is pretty bad (maybe 20% accuracy) and I already used a lot of different functions and roughly 1000 lines of code.

Therefore, I would like to try out a completely new approach with semantic segmentation using deep learning. I looked into a lot of tutorials and all of them seem to use examples with neural networks, which are pretrained for the classes they are detecting the images for (e.g. "cats","dogs" or "airplane"). Obviously the detection works for them, since the neural networks were already trained with a bunch of these images. In my images I would like to detect the objects "pills", "used pills" and "blister".

Now, I want to ask you guys, how I can use my own images to train a neural network in order to segment my image into these classes? I already tried out the Image Labeler App and started manually labeling the images in the different classes. But I don't really know how many images I need and what to do with the the labeled images then.

Please, if anybody can help me in any way I would be very thankful. If you have any doubts that this would work or any other recommendation, please feel free to comment! Thanks in advance!