The digit predictor requires Yolo V2 Region of Interest Detector, as downloadable from this great work, based on a ResNet 50. The model for The predictor is CNN Convnet, able to achieve 0.95 accuracy on cropped images.
- python 3.5
- anaconda 4.4.0
- tensorflow 1.2.1
- keras 2.1.1
- opencv 3.3.0
- imgaug
- Etc.
I recommend to create an anaconda env that is independent of your project. You can create anaconda env for this project by following these simple steps. This process has been verified on MAC OS X10.
$ conda create -n SVHNDetector python=3.5 anaconda=4.4.0
$ source activate SVHNDetector"
(yolo) $ pip install -r requirements.txt
In this project, 2 distinct model will be used:
- a detector for the region of interest, which pretrained weight are
called
detector.h5 - a predictor model on single digit which pretrained weight are called `detector.h5.
Please notice that all the image to be predicted must be in the folder tests\dataset\svhn\imgs; for each call the code
will load both models.
This project provides a way to execute following action from the scratch:
- Download Training and Test data
- Unzip them
- Preprocess them in numpy format containing single images cropped
- start the training test.
Please notice that is a really time consuming operation, spanning for 5-6 hours. Because of this 2 flags can be setted to bypass operations. the first is controlling downloand and unzipping, the second preprocessing.
traintest(False,False)Bypass Download and Preprocessing.
Results are testable with visual plotting, running a jupyter notebok with following command.
jupyter notebook detection_example.ipynb