Using neural network models to classify image objects
Methods, systems, and apparatus, including computer programs encoded on computer storage media, for machine learning. One of the methods includes receiving an image; providing the image to a neural network model, wherein the neural network model is trained to output predictions of one or more locations within the image and corresponding classifications; extracting text content within one or more of the one or more locations; analyzing the extracted text content using the corresponding classifications to evaluate one or more of external consistency with other data records or internal consistency with content from one or more of the particular locations; and generating one or more outputs based on the analyzing.
1 . A method comprising:
receiving an image;
providing the image to a neural network model, wherein the neural network model is trained to output predictions of one or more locations within the image, each location defined by a respective bounding box, and corresponding classifications for each location;
extracting text content within one or more of the bounding boxes output by the neural network model;
analyzing the extracted text content using the corresponding classifications to evaluate one or more of external consistency with other data records or internal consistency with content from at least one other location of the one or more locations; and
generating one or more outputs based on results of the analyzing the extracted text content to evaluate one or more of external consistency or internal consistency.
2 . The method of claim 1 , wherein each location within the image is defined by bounding box coordinates defining a particular region of the image.
3 . The method of claim 2 , wherein the extracting text content comprises performing character recognition within the boundaries of each bounding box.
4 . The method of claim 1 , wherein the neural network model is a convolutional neural network and wherein the convolutional neural network is trained based on a collection of labeled images defining coordinates of regions of interest in the respective documents along with labeled classifications for each region of interest.
5 . The method of claim 4 , wherein the collection of labeled images comprise both human-labeled images and augmented images.
6 . The method of claim 1 , wherein the generating one or more outputs comprises flagging the image as requiring additional evaluation.
7 . The method of claim 1 , wherein each image corresponds to a document, wherein the document is a form having a plurality of fields, each field having a form label and wherein one or more of the fields have been filled out with information responsive to the corresponding form labels.
8 . One or more non-transitory computer storage media encoded with instructions that, when executed by one or more computers, cause the one or more computers to perform operations comprising:
receiving an image;
providing the image to a neural network model, wherein the neural network model is trained to output predictions of one or more locations within the image, each location defined by a respective bounding box, and corresponding classifications for each location;
extracting text content within one or more of the bounding boxes output by the neural network model;
analyzing the extracted text content using the corresponding classifications to evaluate one or more of external consistency with other data records or internal consistency with content from at least one other location of the one or more locations; and
generating one or more outputs based on results of the analyzing the extracted text content to evaluate one or more of external consistency or internal consistency.
9 . The non-transitory computer storage media of claim 8 , wherein each location within the image is defined by bounding box coordinates defining a particular region of the image.
10 . The non-transitory computer storage media of claim 9 , wherein the extracting text content comprises performing character recognition within the boundaries of each bounding box.
11 . The non-transitory computer storage media of claim 8 , wherein the neural network model is a convolutional neural network and wherein the convolutional neural network is trained based on a collection of labeled images defining coordinates of regions of interest in the respective documents along with labeled classifications for each region of interest.
12 . The non-transitory computer storage media of claim 11 , wherein the collection of labeled images comprise both human-labeled images and augmented images.
13 . The non-transitory computer storage media of claim 8 , wherein the generating one or more outputs comprises flagging the image as requiring additional evaluation.
14 . The non-transitory computer storage media of claim 8 , wherein each image corresponds to a document, wherein the document is a form having a plurality of fields, each field having a form label and wherein one or more of the fields have been filled out with information responsive to the corresponding form labels.
15 . A system comprising one or more computers and one or more storage devices storing instructions that are operable, when executed by the one or more computers, to cause the one or more computers to perform operations comprising:
receiving an image;
providing the image to a neural network model, wherein the neural network model is trained to output predictions of one or more locations within the image, each location defined by a respective bounding box, and corresponding classifications for each location;
extracting text content within one or more of the bounding boxes output by the neural network model;
analyzing the extracted text content using the corresponding classifications to evaluate one or more of external consistency with other data records or internal consistency with content from at least one other location of the one or more locations; and
generating one or more outputs based on results of the analyzing the extracted text content to evaluate one or more of external consistency or internal consistency.
16 . The system of claim 15 , wherein each location within the image is defined by bounding box coordinates defining a particular region of the image.
17 . The system of claim 16 , wherein the extracting text content comprises performing character recognition within the boundaries of each bounding box.
18 . The system of claim 15 , wherein the neural network model is a convolutional neural network and wherein the convolutional neural network is trained based on a collection of labeled images defining coordinates of regions of interest in the respective documents along with labeled classifications for each region of interest.
19 . The system of claim 18 , wherein the collection of labeled images comprise both human-labeled images and augmented images.
20 . The system of claim 15 , wherein the generating one or more outputs comprises flagging the image as requiring additional evaluation.
21 . The system of claim 15 , wherein each image corresponds to a document, wherein the document is a form having a plurality of fields, each field having a form label and wherein one or more of the fields have been filled out with information responsive to the corresponding form labels.