This simple/toy project amis to develop an application that adjusts documents after taking pictures. Most of mobile phone are capable of adjusting natural images but not documents (especially the handwritten documents).

The (private) dataset is constituted by

• More than 5,000 pictures of my personal handwritten scribe note since 6 years. (mostly scientific note, including mathematics, physics and chemistry)
• More than 1,000 Resumes
• More than 10,000 pages of mathematic/physic documents/papers (typed)
• More than 500 administrative documents (bill, tax form, contract, etc)
• More than 10,000 natural pictures
• More than 5,000 graffiti, illustration, manga, animation...etc

Developing a network that fits for the dataset with a such diversity isn't simple. By focusing on both handwritten and typed documents, I have achieved 99,93% accuracy for four classes (0, 90, 180, 270 degrees rotation) classification. Note that in the handwritten collection, some pictures are alomost blank with scribbled note. The main idea is to pass a filter (unsupervised) that crops the interesting area then to train the network on patch of images. As a comparison, without these preprocessing, the accuracy is only 70%.

Further improvement could be done by introducing some transfer learning techniques. (TODO)