Neural network architecture for classifying documents

1. A method of classifying a document, comprising:

receiving, from a storage device, an image of the document;

obtaining a predefined user configuration indicative of a term of interest for the document;

converting, by a document importer, the image to machine-readable data using Optical Character Recognition (OCR);

performing, by a first convolutional neural network, semantic enrichment by highlighting the term of interest in the image based on the machine-readable data;

splitting, by a second convolutional neural network, the image into four quadrants for identifying a positional context of the term of interest in the quadrants, wherein the first convolutional neural network and the second convolutional neural network are a ResNet-152 model;

generating a model representation for each of the quadrants;

concatenating the model representations of the quadrants; and

classifying the image based on the concatenated model representations.

2. The method of claim 1 , further comprising:

identifying coordinates of the term of interest from the machine-readable data;

generating a grayscale version of the image based on the identification; and

performing, by the first convolutional neural network, the semantic enrichment of the grayscale version of the image by highlighting the term of interest in the image.

3. The method of claim 1 , further comprising:

converting quadrant images into a shape of 224×224×3; and

generating the model representation for each of the quadrant images for the concatenation and classification.

4. The method of claim 1 , comprising:

generating the model representation from the four quadrants using a fully connected layer; and

classifying the image using a softmax neural network layer.

5. The method of claim 4 , wherein the fully connected layer is trained using a few-shot learning approach.

6. A system for classifying a document, comprising:

a storage device configured to:

store an image derived from a document having text information, and

store a predefined user configuration, indicative of a term of interest for the document;

a document importer configured to:

convert image data in the image to machine-readable data using Optical Character Recognition (OCR); and

a neural network configured to:

perform semantic enrichment by highlighting the term of interest in the image based on the machine-readable data;

split the image into four quadrants for identifying a positional context of the term of interest in the quadrants;

generate a model representation for each of the quadrants;

concatenate the model representations of the quadrants; and

classify the image based on the concatenated model representations,

wherein the neural network comprises a first convolutional neural network to process the machine-readable data for semantic enrichment and a second convolutional neural network to split the image, wherein the first convolutional neural network and the second convolutional neural network are a ResNet-152 model.

7. The system of claim 6 , wherein the document importer is further configured to:

identify coordinates of the term of interest from the machine-readable data; and

generate a grayscale version of the image based on the identification, and

wherein the first convolutional neural network is configured to perform the semantic enrichment of the grayscale version of the image by highlighting the term of interest in the image.

8. The system of claim 6 , wherein the neural network is configured to:

convert quadrant images into a shape of 224×224×3; and

generate the model representation for each of the quadrant images for the concatenation and classification.

9. The system of claim 6 , wherein the neural network further comprises:

a fully connected layer to generate the model representation from the four quadrants; and

a softmax layer to classify the image.

10. The system of claim 9 , wherein the fully connected layer is trained using a few-shot learning approach.