0

I have the jpeg images much wider and higher than values, I'm passing with ImageRecordReader(long height, long width, long channels), say ImageRecordReader(100, 100, 1), while the images themselves are 2000x3000 pixels size and have 3 channels. Will the dl4j process the images vectorization correctly and take not the first 10000 pixel lines (i.e 5 lines o 2000-long lines) but really a square of 100x100 in the top right part of the image?

Eljah
  • 4,188
  • 4
  • 41
  • 85
  • 1
    The images will get resized, so I suppose the answer is "no". If you're looking for how to accomplish that, please reformulate your question. – Samuel Audet Feb 12 '20 at 23:29
  • @SamuelAudet I suppose I know how to accomplish, just manipulate the image data before initializing the ImageRecordReader to crop the needed part from that. So the question is how the images will be resized from 2000x3000 to 100x100 and where are the docs about it? Don't see any valid explanation in javadocs – Eljah Feb 13 '20 at 06:47
  • 1
    The documentation is scattered around, but it's mentioned here, for example: https://deeplearning4j.org/docs/latest/datavec-overview – Samuel Audet Feb 14 '20 at 07:37
  • 1
    @SamuelAudet I've missed that though... But I can see here the default constructor description mentioning rescale, not the constructor explicitly accepting the pixel sizes. – Eljah Feb 14 '20 at 22:22
  • @SamuelAudet could you please post your answer as the official answer that explained the question? – Eljah Oct 06 '20 at 11:53

0 Answers0