My problem statement is as follows: I've got some technical documents as JPEGs, mostly text, but with some standard pictograms. Each image does not need to have all the pictograms or any pictogram at all. I need to find out which pictograms exist in a particular image. I am trying to solve this using Deep Learning.
Further information:
-> The documents are high res, approx 4000 x 6000 px
-> The pictograms are approx. 1/10th the doc size
-> The pictograms belong to 3 classes
There are two approaches I am considering:
1.) Object classification problem: X: image, Y: 3 independent binary classification probabilities as the output of a pretrained CNN (InceptionNet)
2.) Object detection problem: X: image, Y: bounding boxes for the pictograms and the associated class probabilities using a pretrained model (YOLO)
My question is as follows:
-> Is there an advantage to pursuing the Object Detection route? For a given number of training examples,does Object detection learn a more robust function which takes into account the correlation between the spatial information of the boxes and the pictogram classes?
I am a little sceptical as the pictograms are small and surrounded by text. Although, I am considering using Medianblur in OpenCV to reduce some of the background noise.
What would be a better approach?
Thanks in advance.
TLDR; for the number of training examples, does object detection produce a better classifier?
