Systems and methods for real-time end-to-end capturing of ink strokes from video
A real-time end-to-end system for capturing ink strokes written with ordinary pen and paper using a commodity video camera is described. Compare to traditional camera-based approaches, which typically separate out the pen tip localization and pen up/down motion detection, described is a unified approach that integrates these two steps using a deep neural network. Furthermore, the described system does not require manual initialization to locate the pen tip. A preliminary evaluation demonstrates the effectiveness of the described system on handwriting recognition for English and Japanese phrases.
1. A system comprising:
at least one camera for acquiring a video of a handwriting of a user; and
a processing unit for processing the acquired video of the handwriting of the user using a neural network to detect a plurality of ink strokes and for recognizing the handwriting of the user using the detected plurality of ink strokes,
wherein the neural network comprises an encoder and a decoder,
wherein the encoder inputs a sequence of video frames of the acquired video of the handwriting of the user and generates feature maps learned from convolutional network blocks,
wherein the encoder uses a recurrent neural network to generate feature representations based on the feature maps learned from the convolutional network blocks,
wherein the decoder receives the feature maps from the convolutional network blocks and converts the feature representations from the recurrent neural network into pixel-wise label maps using deconvolutional network blocks and the feature maps learned by the decoder from the convolutional network blocks, and
wherein the convolutional network blocks are linked with the deconvolutional network blocks using a plurality of links.
2. The system of claim 1 , wherein the neural network is a deep learning neural network.
3. The system of claim 1 , wherein the camera is a webcam.
4. The system of claim 1 , wherein the camera is mounted above a tabletop.
5. The system of claim 1 , wherein the camera is mounted on a desk lamp.
6. The system of claim 1 , wherein the handwriting of the user is recognized using a handwriting recognition engine.
7. The system of claim 1 , wherein the processing and recognizing are performed in real-time.
8. The system of claim 1 , wherein the processing unit outputs a textual information based on the recognized handwriting of the user.
9. The system of claim 1 , wherein the handwriting of the user comprises ink written with ordinary pen and paper.
10. The system of claim 1 , further comprising using the pixel-wise label maps for consecutive video frames of the acquired video to simultaneously track a pen-tip and classify the consecutive video frames as pen-up or pen-down, wherein the plurality of ink strokes is detected based on classifying the consecutive video frames as pen-up or pen-down.
11. A method comprising:
using at least one camera to acquire a video of a handwriting of a user; and
using a processing unit to process the acquired video of the handwriting of the user using a neural network to detect a plurality of ink strokes and for recognizing the handwriting of the user using the detected plurality of ink strokes,
wherein the neural network comprises an encoder and a decoder,
wherein the encoder inputs a sequence of video frames of the acquired video of the handwriting of the user and generates feature maps learned from convolutional network blocks,
wherein the encoder uses a recurrent neural network to generate feature representations based on the feature maps learned from the convolutional network blocks,
wherein the decoder receives the feature maps from the convolutional network blocks and converts the feature representations from the recurrent neural network into pixel-wise label maps using deconvolutional network blocks and the feature maps learned by the decoder from the convolutional network blocks, and
wherein the convolutional network blocks are linked with the deconvolutional network blocks using a plurality of links.
12. The method of claim 11 , wherein the neural network is a deep learning neural network.
13. The method of claim 11 , wherein the camera is a webcam.
14. The method of claim 11 , wherein the camera is mounted above a tabletop.
15. The method of claim 11 , wherein the camera is mounted on a desk lamp.
16. The method of claim 11 , further comprising using the processing unit to output a textual information based on the recognized handwriting of the user.
17. The method of claim 11 , wherein the handwriting of the user comprises ink written with ordinary pen and paper.
18. A tangible computer-readable medium embodying a set of instructions implementing a method comprising:
using at least one camera to acquire a video of a handwriting of a user; and
using a processing unit for process the acquired video of the handwriting of the user using a neural network to detect a plurality of ink strokes and for recognizing the handwriting of the user using the detected plurality of ink strokes,
wherein the neural network comprises an encoder and a decoder,
wherein the encoder inputs a sequence of video frames of the acquired video of the handwriting of the user and generates feature maps learned from convolutional network blocks,
wherein the encoder uses a recurrent neural network to generate feature representations based on the feature maps learned from the convolutional network blocks,
wherein the decoder receives the feature maps from the convolutional network blocks and converts the feature representations from the recurrent neural network into pixel-wise label maps using deconvolutional network blocks and the feature maps learned by the decoder from the convolutional network blocks, and
wherein the convolutional network blocks are linked with the deconvolutional network blocks using a plurality of links.