The problem of automatic handwritten character recognition (HCR) can be decomposed into two distinct categories, i.e., online HCR and offline HCR. The former one, being much easier to solve than the other, corresponds to the real- time recognition of handwritten characters when the hand strokes of the writer is available to the machine learning system. As the procedural way people write is quite similar, such as the starting point, direction and order of drawing lines, this is rather a solved problem in machine learning: Nowadays, almost every smartphone provides such a feature or application. On the other hand, offline recognition is much complicated as the only information available to the machine learning system is the output of the hand strokes, i.e., an image, which is nothing but a matrix of numbers (corresponding to pixels). In addition to this, the difficulty of offline HCR especially for postal address recognition comes from many aspects such as:
- Complexity: Obviously, the recognition problem is very complex, because every person writes with his / her distinct style. This includes size of letters, italics angle, amount of gap between letters, words and lines etc.
- The Need for Annotated Data: Such a complex problems can only be solved with a supervised machine learning approach
- Cursive Handwritings: It is easier to recognize single unconnected letters/digits. Cursive (connected) handwritings require complex preprocessing and image segmentation algorithms. The recognition performances, unfortunately, depend highly on this process
- Varying Pen Styles: The thickness of the letters vary from pen to pen even for the same individual
- Size of images: Image sizes are different
- Artifacts: There are artifacts acting as noise in some images. These can be logos, lines, marks etc.
- Direction of Lines: As there are no strict horizontal lines, some people tend to write in a tilted way
Syslore Multiline Handwritten OCR is based on the cutting-edge deep learning research solving all above mentioned challenges. It adopts the strong aspects of Stacked Denoising Autoencoders, Deep Convolutional Networks and Recurrent Neural Networks and combines them to maximising the performance. Syslore Multiline Handwritten OCR combines state-of-the art deep learning based Region of Interest detection engine (Syslore Droidi) and extremely high performance fuzzy address lookup engine (Syslore Match) yielding outstanding read rates being capable of reading from the postal code level all the way to delivery point and recipient level with very low error rates making it a perfect solution for postal operators having a need for maximum recognition rates for handwritten mail pieces.
Commercially Proven Technology
As a CEN compliant solution, Syslore Multiline Handwritten OCR can be flexibly utilized with other solutions, as a primary OCR or as a secondary address reader to further improve the performance of existing address reading and mail sorting solutions reducing manual mail sorting and video coding work significantly. Therefore multiple postal operators world wide rely on Syslore Multiline Handwritten OCR.