I am given a task to locate the defect in images. However, they have just given only 1 sample each for a good part and bad part. I am not understanding how to make the dataset for this. 
The image with 2 plain holes is bad part where as the other is a good part.
Are you sure you have only a single input image for both classes without any validation and test set?
How will your model be evaulated?
The task seems to be a bit “weird”.
This task for sure is a bit “weird”. There should be no reason why you can’t get enough images, it just a matter of few more camera clicks, in fact you record a video and then extract frame by frame images to create your dataset.
In any case all above is not possible then there is only one method left, you can use massive data augmentation on that single image. More about data transformation here.
What he means is that he has only 1 image for the image segmentation task. Perhaps OP wants you could point to any source where more images can be obtained