OverView:Using FCN to segment the book's content and background, then dewarping the pages.
Last Updated Code:2018.06.19 Continuing......
First Step:
Using FCN(fully convolution network) to segment the image into 3 parts(left page,right page and background).
DataSet Created By Lin YangBin Now we have 500 images and labeled images for book pages.
Using GTX 1070 8GB Trained this network for 8 hours. loss value ->0.01
TODO:Data Augment and Dewarp Algorithm.
1.You should download the trained VGG model from http://www.vlfeat.org/matconvnet/models/beta16/imagenet-vgg-verydeep-19.mat, put this model into the folder "./pre_data".
2.Prepared the dataset for training and validation.
3.cmd->cd to your path->python main,py->running(first training your model use this network, then change the flag.mode to "test" mode or "visual" mode to identify the results).
loss value
https://www.cv-foundation.org/openaccess/content_cvpr_2015/html/Long_Fully_Convolutional_Networks_2015_CVPR_paper.html https://github.com/shelhamer/fcn.berkeleyvision.org https://github.com/shekkizh/FCN.tensorflow