HANDWRITING RECOGNITION WITH CONVOLUTIONAL NEURAL NETWORK GOOGLENET ON MATLAB
As technology becomes more advanced, human life becomes much easier, although humans still use pen and paper to write data that they need to keep. This becomes a problem when written text need to be changed into digital text with typing each letter one by one, where it takes too much time and ineffi...
Saved in:
| Main Author: | |
|---|---|
| Format: | Final Project |
| Language: | Indonesia |
| Online Access: | https://digilib.itb.ac.id/gdl/view/62312 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Institution: | Institut Teknologi Bandung |
| Language: | Indonesia |
| Summary: | As technology becomes more advanced, human life becomes much easier, although humans still use pen and paper to write data that they need to keep. This becomes a problem when written text need to be changed into digital text with typing each letter one by one, where it takes too much time and inefficient. Optical Character Recognition (OCR) can be a solution to answer this problem, whereas this study strives to make machine recognize images like humans do. With Convolutional Neural Network (CNN) as its brain, features of a character image will be extracted automatically, making it a practical process and preventing it to overload computational strength with its architecture.
In this study, 650 images of writer handwritten text are provided as training and validation data. Preprocessing will be held to optimize data that will be used in CNN training phase, such as thresholding, noise reduction, and character segmentation. The trained GoogLeNet has validation accuracy of 98,9% and the training duration was 3 minutes and 15 seconds. The validity of the trained network will be held with character recognition test and has 96,15% accuracy. Line segmentation, word segmentation, and character segmentation completes the final prototype. The prototype successfully displays the network reading result of a tested handwritten text image. |
|---|