I am currently working on a defect detection algorithm but I only have a few samples of defects.I googled for defect detection datasets and I found this one:


which has a few hundreds of original images of defects.

My idea is: Imagenet => Defect dataset from internet => Own defect dataset

Step 1. Training a model with ImageNet initialization using the defect dataset found in the internet (+ non-defect images + augmented data)

Step 2. Using the output model of step 1 (which will be more similar to my own data),do transfer learning using my own defect dataset (defects + non-defects + augmented).

Do you think this a good way to get good results?

Based on: https://blog.slavv.com/a-gentle-intro-to-transfer-learning-2c0b674375a0

Should defect images consider as low similar with imagenet's images? or similar to model because a both inputs are images? Some webpages said because they both are images, they are similar but some webpages said because these images are too different to the images used to train the imagenet model so I got confused about this.

If I skip step 1, I dont think I get anything good because I have less than 100 images.

Any advise or comment will be appreciated.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.