Method and apparatus for classifying machine printed text and handwritten text

A non-transitory computer-readable medium storing a plurality of instructions which, when executed by a processor, cause the processor to perform operations for classifying a machine printed text and a handwritten text in an input, the operations comprising:

• defining a perspective for an auto-encoder, wherein the perspective comprises an inpaintinq perspective, wherein the inpainting perspective is a function of a hidden representation of the input and a reconstruction of the input, wherein the hidden representation is based on an encoding function comprising a first non-linear function of an encoding layer weight of the input plus an encoding layer bias, wherein the reconstruction of the input is based on a decoding function comprising a second non-linear function of a decoding layer weight of the input plus a decoding layer bias;
• receiving the input for the auto-encoder, wherein the input comprises a document comprising the machine printed text and the handwritten text;
• performing an encoding on the input using an auto-encoder to generate a classifier;
• applying the classifier on the input;
• and
• generating an output that classifies the machine printed text and the handwritten text in the input based on the classifier in accordance with the perspective.

A method for classifying a machine printed text and a handwritten text in an input, comprising:

• defining, by a processor, a perspective for an auto-encoder, wherein the perspective comprises an inpainting perspective that trains two models, the two models comprising a first model for the machine printed text and a second model for the handwritten text, wherein the inpaintinq perspective is a function of a hidden representation of the input and a reconstruction of the input, wherein the hidden representation is based on an encoding function comprising a first non-linear function of an encoding layer weight of the input plus an encoding layer bias, wherein the reconstruction of the input is based on a decoding function comprising a second non-linear function of a decoding layer weight of the input plus a decoding layer bias, or a classification perspective that trains a single model, wherein the classification perspective is a modification of the inpaintinq perspective that changes output units to softmax units;
• receiving, by the processor, the input for the auto-encoder, wherein the input comprises a document comprising an unknown mixture of the machine printed text and the handwritten text, wherein the auto-encoder has not been previously trained with a training model;
• performing, by the processor, an encoding on the input using an auto-encoder to generate a classifier via a sliding window approach;
• applying, by the processor, the classifier on the input via the sliding window approach;
• generating, by the processor, an output, wherein the output comprises either two outputs for the inpainting perspective or a single output for the classification perspective that classifies the machine printed text and the handwritten text in the input based on the classifier;
• and
• forwarding, by the processor, the output that has the machine printed text and the handwritten text classified to an optical character reader or an intelligent character reader.