I am working on a text extraction problem from Invoices. I want to detect various fields in the invoice like the following.
I am struggling to find any dataset for invoices. I have a dataset of 150 images of invoices of 4-5 different templates. I trained a pretrained Resnet 50 using tensorflow object detection API for 4000 epochs. To evaluate results i have a test set of just 10 images. The model is not doing well on test set. It has approx 0.66 mAP on test set. And its doing good on train set but not perfect. I get approx 0.94 mAP.
I want to make model work perfect on my Test set. So next time the model sees a new invoice (similar to train set) it should predict bounding boxes with a very high accuracy. How should I do this. I have a very small dataset. I cannot gather more than 300-400 images for invoices.
I found a large dataset of Reciepts. I think I can use that to fine tune Resnet 50. And then again Fine tune the model for my invoices dataset. Will this approach work???