Embedded Files
DESIGN
Braille Cell
In the design of a single Braille cell, six pins are affixed to a surface, each connected to springs positioned above a slender solenoid. The solenoid is created by a few hundred turns of a 36 gauge copper wire with an enamel coating over an iron core. Upon the passage of an input current through the solenoid, a magnetic field is generated. The pin is attracted towards the solenoid, resulting in a downward vertical displacement. The pin hence sinks down inside the surface. Upon deactivation of the current, the pin returns to its initial position through spring force. Consequently, this coordinated activation and deactivation of the six solenoids associated with each of these six pins results in the formation of a tactile pattern corresponding to a single Braille alphabet, facilitating touch-based recognition
Braille Cell
OPTICAL CHARACTER RECOGNITION (OCR)
We implemented Vanilla Transformer encoder-decoder model architecture for optical character recognition. The encoder takes an image as input and generates its visual representations. The decoder generates wordpiece sequence based on the visual features and previous prediction.
Encoder: The encoder receives a colored image(RGB format) of dimension (H0×W0x3) as input, which is then resized to a standard dimension(H, W). As Transformer encoder requires input data as sequences of tokens, the encoder undertakes a decomposition of the input image. The image is decomposed into a batch of N = H*W/P^2 square patches, each having a fixed size of (P, P). These batches are passed through the encoder to generate visual representations.
Decoder: We employed the conventional Transformer decoder, which consists of a series of identical layers, mirroring the structure found in the encoder. However, there is an inclusion of "encoder-decoder attention" between he multi-head self-attention and feed-forward network components. This introduced element serves the purpose of directing distinct attention mechanisms towards the encoder’s output. To maintain a controlled flow of information during training and prediction, the decoder incorporates attention masking in its self-attention mechanism. The decoder’s hidden states are subject to linear trans- formation, wherein they are projected from the model’s dimension to match the vocabulary size denoted as ’V.’ Subsequently, probabilities pertaining to the vocabulary elements are computed using the softmax function applied to this transformed representation
PROTOTYPE (1 Braille Cell)
The Braille cells are equipped with solenoid and spring mechanisms to raise and lower the pins accurately. OCR functionality has been successfully integrated into the device, utilizing the TrOCR engine for accurate text recognition. Users have the capability to upload images of printed text using their laptops or PCs and convert them into Braille for tactile reading. Through careful component selection and efficient design, the production costs have been kept down, resulting in a device that is significantly more affordable than existing Braille displays with similar capabilities.
Page updated
Google Sites
Report abuse