We claim:
1. A system (100) for efficiently identifying a subject, the system (100) comprising:
a memory (126) storing program instructions;
a processor (124) configured to execute instructions stored in the memory (126); and
an identification engine (122) executed by the processor (124) and configured to:
segment micro-voltage digital signals into intervals of a pre-defined time period;
transform the segmented micro-voltage digital signals into a frequency domain for computing on a Mel’s scale, wherein the Mel’s scale provides a unique signature of the subject in the form of a Melspectrogram image; and
pass the Melspectrogram image through a trained deep learning model, wherein features associated with the Melspectrogram image are extracted into a feature map for obtaining predicted labels associated with the subject based on labels used during training of the deep learning model for identifying the subject.
2. The system (100) as claimed in claim 1, wherein the intervals of the pre-defined time period comprises ±10 seconds or less.
3. The system (100) as claimed in claim 1, wherein the identification engine (122) comprises a computation unit (128) executed by the processor (124) and configured to receive the micro-voltage digital signals from a data receiver unit (118) and process the micro-voltage digital signals for segmenting into intervals of the pre-defined time period.
4. The system (100) as claimed in claim 1, wherein the identification engine (122) comprises a prediction unit (130) executed by the processor (124) and configured to generate the deep learning model using neural networks associated with the deep learning techniques, and wherein the deep learning techniques comprise a Deep Neural Network (DNN), a Long Short Term Memory Network (LSTM) and a Convolutional Neural Network (CNN).
5. The system (100) as claimed in claim 1, wherein one or more pre-defined number of neural network layers of the deep learning model are stacked together through which the Melspectrogram image is passed, and wherein the pre-defined number of neural network layers of the deep learning model comprises three convolution 2-D layers paired with three max pooling 2-D layers respectively, two dense layers, a flattening layer between the two dense layer and a dropout layer.
6. The system (100) as claimed in claim 5, wherein the three convolution 2-D layers paired with the max pooling 2-D layers are used for extracting features from the Melspectrogram images and subsequently carrying out downsampling, and wherein the features are associated with the subject.
7. The system (100) as claimed in claim 5, wherein the prediction unit (130) passes a 1-D tensor associated with the Melspectrogram images through the flattening layer to the dense layer, and wherein the dropout layer between the two dense layers prevents the deep learning model from over fitting.
8. The system (100) as claimed in claim 1, wherein the deep learning model is trained in a prediction unit (130) of the identification engine (122) with training datasets prior to identifying the subject, the training datasets are generated based on captured micro-vibrations associated with multiple subjects in a resting position and converting the captured micro-vibrations into the Melspectrogram image.
9. The system (100) as claimed in claim 8, wherein the training datasets are pre-processed and inputted to the prediction unit (130) of the identification engine (122) along with labels for training the deep learning model, the labels used in training the deep learning model represent ground truth associated with each Melspectrogram image, and wherein the training datasets comprises input images associated with the multiple subjects, the input images are Melspectrogram images in a 4-D format with dimensions “batch_size, height, width, depth”, such that the batch_size is a number of training datasets in one forward pass, height (H) is height of the image, width (W) is width of the image, and depth (D) is number of color channels of the image.
10. The system (100) as claimed in claim 9, wherein the prediction unit (130) trains the deep learning model by passing Melspectrogram images through the convolution layers and max pooling 2-D layers along with the respective labels, and wherein the batch_size of the output images remains same as that of the input Melspectrogram images and height, weight and depth of the output image changes based on number of filters, kernels and padding of the convolution layers.
11. The system (100) as claimed in claim 10, wherein the filters of the convolution layer of the deep learning model comprises light color regions and dark color regions such that the light color region in the filter represents a value ‘1’, and the dark color region in the filter represents a value ‘0’.
12. The system (100) as claimed in claim 10, wherein the prediction unit (130) passes the Melspectrogram image through the three convolution 2-D layers paired with three max pooling 2-D layers by providing the Melspectrogram image as an input to a first convolution layer and output of the first convolution layer is provided as an input to a first max pooling 2-D layer, the output of the first max pooling 2-D layer is provided as an input to a second convolution layer and output of the second convolution layer is provided as an input to a second max pooling 2-D layer, the output of the second max pooling 2-D layer is provided as an input to a third convolution layer and output of the third convolution layer is provided as an input to a third max pooling 2-D layer.
13. The system (100) as claimed in claim 12, wherein the prediction unit (130) passes the Melspectrogram image through the trained deep learning model for computing the identity of the subject by:
pre-processing the Melspectrogram image associated with the subject to compute a Melspectrogram image of dimensions (None, 32, 32, 3), the dimension ‘None’ represents various numbers of images which are provided while training and “32, 32, and 3” represents height (H1), width (W1) and depth (D1) respectively of the Melspectrogram image;
providing the computed Melspectrogram image as an input to the first convolution layer of the deep learning model to generate an output of dimension “None, 30, 30, 16” based on number of filters (K1) in the first convolution layer, strides (S), spatial extent of the filters (F) and padding (p), wherein the number of filters is 16, strides (S) is 1 and spatial extent of filters (F) is 3 with 0 padding (P);
providing the output from the first convolution layer as an input to a first max pooling 2-D layer, the first max pooling 2-D layer uses a shape of dimensions (2, 2) for reducing dimensions of the output received from the first convolution layer to generate an output of a dimension “None, 15, 15, 16”, wherein ‘None’ represents various numbers of images which are provided while training and “15, 15, 16” represents height, width and depth of the output;
providing the output from the first max pooling 2-D layer as an input to the second convolution layer to generate an output of a dimension “None, 13, 13, 16”, wherein ‘None’ represents various numbers of images which are provided while training and “13, 13, 16” represents height (H2), width (W2) and depth (D2) of the output;
providing the output from the second convolution layer to as an input to a second max pooling 2-D layer to generate an output of a dimension “None, 6, 6, 16”, wherein ‘None’ represents various numbers of images which are provided while training and “6, 6, 16” represents height, width and depth of the output;
providing the output from the second max pooling 2-D layer to a third convolution layer to generate an output of a dimension “None, 4, 4, 16”, wherein ‘None’ represents various numbers of images which are provided while training and “4, 4, 16” represents height (H3), width (W3) and depth (D3) of the output; providing the output from the third convolution layer as an input to a third max pooling 2-D layer to generate an output of a dimension “None, 2, 2, 16”, wherein ‘None’ represents various numbers of images which are provided while training and “2, 2, 16” represents height, width and depth of the output;
providing the output from the third max pooling 2-D layer as an input to a flattening layer to generate an output of a dimension “None, 64”;
providing the output from the flattening layer as an input to a first dense layer to generate an output of a dimension “None, 256”;
providing the output from the first dense layer as an input to a dropout layer to generate an output of a dimension “None, 256”;
providing the output from the dropout layer as an input to a second dense layer to generate an output of a dimension “None, 20”, wherein the value “20” represents number of the labels; and
transforming the output from the second dense layer into a feature map using the convolution layer filters present in the convolution layer for obtaining the predicted labels, for identifying the subject, based on the labels used during training of the deep learning model.
14. The system (100) as claimed in claim 13, wherein the output from the second dense layer is associated with the predicted labels used for identifying the subject, and wherein the predicted labels represent response of the trained deep learning model for classifying the Melspectrogram image associated with the subject.
15. The system (100) as claimed in claim 13, wherein the prediction unit (130) is configured to compute parameters based on the first convolution layer, the second convolution layer, the third convolution layer, the first dense layer and the second dense layer of the trained deep learning model, and wherein the parameters represent the number of learnable elements in a convolution layer, and wherein the number of parameters are computed based on the number of filters (K) used along with their kernel size (KZ), a bias and number of filters in the previous layer (D).
16. The system (100) as claimed in claim 15, wherein the parameters are computed during the training of the deep learning model, prior to the implementation of the trained deep learning model.
15.The system (100) as claimed in claim 13, wherein the prediction unit (130) is configured to transmit identity data of the subject associated with the computed identity of the subject to a database (132) in the identification engine (122) for storage and future retrieval, and wherein a user device (134) is configured to connect to the database (132) for retrieving, accessing and viewing the subject’s identity data via a Graphical User Interface (GUI) of an application in the user device (134) or via a GUI rendered via a web portal.
16. The system (100) as claimed in claim 1, wherein the system (100) is configured to compute health data of the subject based on identity data associated with the identity of the subject, and wherein the health data of a non-intended subject captured and tagged intentionally or unintentionally along with the health data of the intended subject is removed based on the identity data of the identified subject, and wherein the health data captured from an intended subject and a non-intended subject is distinguished for preventing mixing of the health data of the intended subject and the non-intended subject.
17. The system (100) as claimed in claim 1, wherein the system (100) is configured to couple identity data associated with the identified subject with the subject’s biometric data for providing double layer secure authentication, wherein the subject’s biometric data comprises retina scan and fingerprints.
18. A method for efficiently identifying a subject, wherein the method is implemented by a processor (124) executing instructions stored in a memory (126), the method comprises:
segmenting micro-voltage digital signals into intervals of a pre-defined time period;
transforming the segmented micro-voltage digital signals into a frequency domain for computing on a Mel’s scale, wherein the Mel’s scale provides a unique signature of the subject in the form of a Melspectrogram image; and
passing the Melspectrogram image through a trained deep learning model, wherein features associated with the Melspectrogram image are extracted into a feature map for obtaining predicted labels associated with the subject based on labels used during training of the deep learning model for identifying the subject.
19. The method as claimed in claim 18, wherein the intervals of the pre-defined time period comprises ±10 seconds.
20. The method as claimed in claim 18, wherein the deep learning model is generated using neural networks associated with the deep learning techniques, and wherein the deep learning techniques comprises a Deep Neural Network (DNN), a Long Short Term Memory Network (LSTM) and a Convolutional Neural Network (CNN).
21. The method as claimed in claim 18, wherein one or more pre-defined number of neural network layers of the deep learning model are stacked together, through which the Melspectrogram image is passed, and wherein the predefined number of neural network layers of the deep learning model comprises three convolution 2-D layers paired with three max pooling 2-D layers respectively, two dense layers, a flattening layer between the two dense layer and a dropout layer.
22. The method as claimed in claim 21, wherein the three convolution 2-D layers paired with the max pooling 2-D layers are used for extracting features from the Melspectrogram images and subsequently carrying out downsampling, and wherein the features are associated with the subject.
23. The method as claimed in claim 18, wherein a 1-D tensor associated with the Melspectrogram images is passed through the flattening layer to the dense layer, and wherein the dropout layer between two dense layers prevents the deep learning model from over fitting.
24. The method as claimed in claim 18, wherein the deep learning model is trained with training datasets prior to identifying the subject, the training datasets are generated based on capturing the subject’s micro-vibrations associated with the multiple subjects in a resting position and converting the captured micro-vibrations into the Melspectrogram image.
25. The method as claimed in claim 24, wherein the training datasets are pre-processed and inputted along with labels for training the deep learning model, the labels used in training of the deep learning model represent ground truth associated with every Melspectrogram image, and wherein the training datasets comprises input images associated with multiple subjects, the input images are Melspectrogram images in a 4-D format with dimensions “batch_size, height, width, depth”, such that the batch_size is a number of training datasets in one forward pass; height (H) is height of the image; width (W) is width of the image; and depth (D) is number of color channels of the image.
26. The method as claimed in claim 22, wherein the Melspectrogram images are passed through the convolution layers and max pooling 2-D layers of the deep learning model along with the respective labels, and wherein the batch_size of the output image remains same as that of input Melspectrogram image and height, weight and depth of the output image changes based on number of filters, kernels and padding of the convolution layers.
27. The method as claimed in claim 21, wherein the Melspectrogram image is passed through the three convolution 2-D layers paired with three max pooling 2-D layers by providing the Melspectrogram image as an input to a first convolution layer and output of the first convolution layer is provided as an input to a first max pooling 2-D layer, the output of the first max pooling 2D layer is provided as an input to a second convolution layer and output of the second convolution layer is provided as an input to a second max pooling 2-D layer, the output of the second max pooling 2-D layer is provided as an input to a third convolution layer and output of the third convolution layer is provided as an input to a third max pooling 2-D layer.
28. The method as claimed in claim 27, wherein the Melspectrogram image is passed through the trained deep learning model for computing the identity of the subject by:
pre-processing the Melspectrogram image associated with the subject to compute a Melspectrogram image of dimensions (None, 32, 32, 3), the dimension ‘None’ represents various numbers of images which are provided while training and “32, 32, and 3” represents height (H1), width (W1) and depth (D1) respectively of the Melspectrogram image;
providing the generated Melspectrogram image as an input to the first convolution layer of the deep learning model to generate an output image of dimension “None, 30, 30, 16” based on number of filters (K1) in the first convolution layer, strides (S), spatial extent of the filters (F) and padding (p), wherein the number of filters is 16, strides (S) is 1 and spatial extent of filters (F) is 3 with 0 padding (P);
providing the output from the first convolution layer as an input to a first max pooling 2-D layer, the first max pooling 2-D layer uses a shape of dimensions (2, 2) for reducing dimensions of the output image received from the first convolution layer to generate an output of a dimension “None, 15, 15, 16”, wherein ‘None’ represents various numbers of images which are provided while training and “15, 15, 16” represents height, width and depth of the output;
providing the output from the first max pooling 2-D layer as an input to the second convolution layer to generate an output of a dimension “None, 13, 13, 16”, wherein ‘None’ represents various numbers of images which are provided while training and “13, 13, 16” represents height (H2), width (W2) and depth (D2) of the output image;
providing the output from the second convolution layer to as an input to a second max pooling 2-D layer to generate an output of a dimension “None, 6, 6, 16”, wherein ‘None’ represents various numbers of images which are provided while training and “6, 6, 16” represents height, width and depth of the output;
providing the output from the second max pooling 2-D layer to a third convolution layer to generate an output of a dimension “None, 4, 4, 16”, wherein ‘None’ represents various numbers of images which are provided while training and “4, 4, 16” represents height (H3), width (W3) and depth (D3) of the output;
providing the output from the third convolution layer as an input to a third max pooling 2-D layer as an input to generate an output of a dimension “None, 2, 2, 16”, wherein ‘None’ represents various numbers of images which are provided while training and “2, 2, 16” represents height, width and depth of the output;
providing the output from the third max pooling 2-D layer as an input to a flattening layer to generate an output of a dimension “None, 64”; providing the output from the flattening layer as an input to a first dense layer to generate an output of a dimension “None, 256”;
providing the output from the first dense layer as an input image to a dropout layer to generate an output of a dimension “None, 256”;
providing the output from the dropout layer as an input
to a second dense layer to generate an output of a
dimension “None, 20”, wherein the value “20” represents number of labels; and
transforming the output from the second dense layer into a feature map using the convolution layer filters present in the convolution layer for obtaining the predicted labels, for identifying the subject, based on the labels used during training of the deep learning model.
29. The method as claimed in claim 28, wherein the output from the second dense layer is associated with the predicted labels used for identifying the subject, and wherein the predicted labels represent response of the trained deep learning model for classifying the Melspectrogram image associated with the subject.