Tensorflow - What is the best model to perform real-grade training?

Question

I’m developing a college project to carry out recognition of the real notes through Tensorflow, but the big problem is that the real notes have a lot of similarity between them, mainly on the back of the ballot, another problem is the lack of a dataset with good quality of the new real ballots, so it is necessary to create a dataset with good quality images and it may be that at this point I am being flawed.

I performed the configuration of all Tensorflow in windows and followed the tutorial (https://github.com/EdjeElectronics/TensorFlow-Object-Detection-API-Tutorial-Train-Multiple-Objects-Windows-10) and there it uses the model Faster-RCNN-Inception-V2-COCO and with the images I have using this model the values of Loss is very high (Very High Even) I wonder if you know any other models that would be very good for my scenario and if you would have some tips around these similarity problems.

Answer 1

Assuming you followed the tutorial correctly, the problem may come from two places:

1. Data: few data, incomplete data, unbalanced dataset
2. Hyperparameters

Let’s start with data: networks deep Learning They need a lot of data, but they’re a lot of data, and that’s why they’re trendy, we finally have to do with all of our data, you said that there are no good real grade datasets, so go to an ATM, get some notes, exchange at the gas station for the 2 and 5, Take your phone and start shooting, remember to use the same lighting and background for different types of note, so the network learns to ignore this information. Or, if you need an easier solution: data augmentation, consists of techniques to generate new data from the data you already have: invert the image, move it a few pixels left or right, cropping etc...

Hyperparameters: these are the fine adjustments of your model, Learning rate, and other parameters of your optimizer, sometimes moving a little in them can help the model to converge faster

I worked a little with Faster-CNN and I can say that it is very good, it seems that the tutorial is doing transfer Learning, retraeing only the last layers of the network, but if you are not going to do it in real time (procesando a video for example) da para usar uma rede pouco menos potente

Answer 2

For pattern recognition could utilize convolutional neural network (CNN) - Vggnet architecture based on some good neural network models, as the CAFFE model recognizes the standard : https://cocodataset.org/#home Abstracting colors and objects like watermark in the cedars.