Granted Patent B2

US 12,401,422 · App. 18/253,781 · Granted Aug 26, 2025

Inventor: Maris Kronbergs (Riga, LV)

H04B10/116G06N3/045G06N3/08G06N3/084H04B10/114H04B10/69H04N19/169

Abstract

Claims (32)

1. A light signal decoding device ( 1 ), comprising:

an image sensor ( 2 ) configured to capture a three different wavelength light signal as an RGB colour light signal,

a memory ( 3 ) configured to store an ANN (Artificial Neural Network) model, wherein the ANN model is a fully connected ANN model with ReLU and Leaky ReLU activation functions, and wherein the ANN model comprises at least three input neurons, two hidden layers and at least two outputs as at least two data bits, and

a processor ( 4 ) configured to:

transform the captured RGB colour light signal into three numeric values;

sum three numeric values to obtain a C (Clear colour) value;

apply the Leaky ReLU activation function to the hidden layers and apply the ReLU activation function to the two output neurons; and

decode, by use of the ANN model, three numeric values and C value into at least two probability values as two output neurons, wherein each probability value is in the range of 0 to 1, and wherein the output bit is 0 if the probability value is less than 0.5, and the output bit is 1 if the probability value is equal to or above 0.5, resulting in two output data bits.

2. The light signal decoding device ( 1 ) according to claim 1 , wherein the input neurons are in the range of 3 to 64 neurons, preferably 3 to 32 neurons, more preferably 3 to 16 neurons, wherein each hidden layer of the ANN model includes at least 10 to 128 nodes, preferably 12 to 64 nodes, more preferably 12 to 32 nodes, and wherein the output neurons are in the range of 2 to 64 neurons, preferably 2 to 32 neurons, more preferably 2 to 16 neurons.

3. The light signal decoding device ( 1 ) according to claim 1 , wherein the ANN model comprises four input neurons, two hidden layers and three output neurons, wherein each input neuron is R colour value, G colour value, B colour value and C colour value of the pixel RGB colour value, wherein each hidden layer of the ANN model includes at least 10 to 128 nodes, preferably 12 to 64 nodes, more preferably 12 to 32 nodes, and wherein at least three output neurons are at least three output data bits as a result of decoded data by means of processor ( 4 ) using the ANN model.

4. Training method of ANN model according to claim 1 , wherein the ANN model is trained using Stochastic Gradient Descent training with augmented training data the training method comprises the steps of:

i) obtaining a training data, wherein the training data comprises of plurality of training numeric values and C value, wherein for each training group, comprising of numeric values and C value, are designated at least two output data bits; and

ii) training the ANN model with the obtained training data to produce a trained ANN model that is capable of producing output data bits based on numeric values and C value, wherein the ANN model is being trained based on the measured output data bits associated with each training group, comprising of numeric values and C value, and wherein the training data comprise varying brightness and colour-biased light signal values.

5. A light signal decoding method, wherein the method comprises the following steps:

a) capturing an RGB colour light signal;

b) transforming the captured RGB colour light signal into three numeric values;

c) summing three numeric values to obtain a C (Clear colour) value;

d) decoding said numeric values and the C value, wherein the step of decoding includes the following sub-steps:

d1) providing each numeric value and the C value as each input neuron to an ANN model;

d2) applying the Leaky ReLU activation function to the hidden layers and applying the ReLU activation function to the three output neurons; and

d3) processing numeric values and the C value as four input neurons through two hidden layers of the ANN model, in result of which three probability values as three output neurons are obtained, wherein each probability value is in the range of 0 to 1, and wherein the output bit is 0 if the probability value is less than 0.5, and the output bit is 1 if the probability value is equal to or above 0.5, resulting in three output data bits.

6. The light signal decoding method according to claim 5 , wherein the step of capturing includes a capture of RGB colour image frame as the RGB colour light signal; and

wherein the step of transforming captured RGB colour image frame further comprises the following steps:

b1) scaling the captured RGB colour image frame to a 2×2 pixel image frame;

b2) calculating an average blue colour value between a top two pixels and a bottom two pixels of the 2×2 pixel image frame;

b3) comparing whether the average blue colour value of the top two pixels of the 2×2 pixel image frame differs from the average blue colour value of the bottom two pixels of the 2×2 pixel image frame;

b31) when the average blue colour value of the top two pixels does not differ from the average blue colour value of the bottom two pixels, then the 2×2 pixel image frame is scaled to 1 pixel image frame;

b32) when the average blue colour value of the top two pixels does differ from the average blue colour value of the bottom two pixels, then the two pixels where the average blue colour value is different from the average blue colour value of the previously sampled two pixels are scaled to 1 pixel image frame and the pixel RGB colour value of the 1 pixel image frame is sent to the ANN model for performing the step c);

b4) comparing whether the average blue colour value of the 1 pixel image frame differs from the average blue colour value of the previously sampled the 1 pixel image frame;

b41) when the average blue colour value of the 1 pixel image frame does differ from the average blue colour value of the previously sampled the 1 pixel image frame, then the pixel RGB colour value of the 1 pixel image frame is sent to the ANN model for performing step c);

b42) when the average blue colour value of the 1 pixel image frame does not differ from the average blue colour value of the previously sampled the 1 pixel image frame, then the 1 pixel image frame is discarded; and

wherein the average blue values sent to the ANN model are stored for sampling with the next average blue values.

Assignments (1)

Priority Claims (1)

Continuity (1)

References Cited (11)