# US 10,121,072 B1 · Unsupervised removal of text from form images > Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt - Number: US 10,121,072 B1 - Document ID: US10121072B1 - Status: Granted - Kind: B1 - Application number: 15/395,728 - Filed: 2016-12-30 - Granted: 2018-11-06 - Earliest claimed: 2016-12-30 - Assignees: Intuit Inc. - Applicants: Intuit Inc. - Inventors: Rohan Kekatpure - Web page: https://patentgrep.com/patent/US10121072B1 ## CPC - G06V 30/412: Layout analysis of documents structured with printed lines or input boxes, e.g. business forms or tables - G06F 40/186: Templates - G06V 10/273: {removing elements interfering with the pattern to be recognised} - G06V 10/431: {Frequency domain transformation; Autocorrelation} - G06V 30/414: Extracting the geometrical structure, e.g. layout tree; Block segmentation, e.g. bounding boxes for graphics or text - G06V: IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING - G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N) ## Abstract The present disclosure relates to language agnostic unsupervised removal of text from form images. According to one embodiment, a method comprises generating a spectral domain representation of an image by applying a two dimensional frequency domain transformation, where the image depicts form layout elements and text elements. Applying a first filter to the spectral domain representation to remove a portion of the frequency domain corresponding to the text element, and applying an inverse two dimensional frequency domain transformation to the filtered spectral domain representation of the image to generate a reconstructed image. The text elements are not depicted in the reconstructed image. ## Claims ### Claim 1 A method for removing text from form images comprising: generating a frequency domain representation of an image by applying a two dimensional frequency domain transformation to the image, wherein the image depicts at least a portion of a form and wherein the form includes layout elements and text elements; removing a portion of the frequency domain representation that represents the text elements by applying a first filter to the frequency domain representation, producing a first filtered frequency domain representation; generating a reconstructed image of at least the portion of the form by applying an inverse two dimensional frequency domain transformation to the first filtered frequency domain representation, wherein the reconstructed image does not depict the text elements; removing a portion of the frequency domain representation that represents the layout elements by applying a layout filler to the frequency domain representation, producing a second filtered frequency domain representation; generating a reconstructed text image by applying an inverse two dimensional frequency domain transformation to the second filtered frequency domain representation, wherein the reconstructed text image does not depict the layout elements; matching the reconstructed image to a form template; generating text from the reconstructed text image; and mapping a portion of the text to a corresponding field of the form template based on a position of the portion of the text in the reconstructed text image. ### Claim 2 (depends on 1) The method of claim 1, wherein the reconstructed image is matched to the form template using a neural network. ### Claim 3 (depends on 1) The method of claim 1, wherein the layout filter comprises a two dimensional high-pass filter. ### Claim 4 A system, comprising: a processor; and memory storing instructions which, when executed on the processor, operate to remove text from form images, the operation comprising: generating a frequency domain representation of an image by applying a two dimensional frequency domain transformation to the image, wherein the image depicts at least a portion of a form and wherein the form includes layout elements and text elements; removing a portion of the frequency domain representation that represents the text elements by applying a first filter to the frequency domain representation, producing a first filtered frequency domain representation; generating a reconstructed image of at least the portion of the form by applying an inverse two dimensional frequency domain transformation to the first filtered frequency domain representation, wherein the reconstructed image does not depict the text elements; removing a portion of the frequency domain representation that represents the layout elements by applying a layout filter to the frequency domain representation, producing a second filtered frequency domain representation; generating a reconstructed text image by applying an inverse two dimensional frequency domain transformation to the second filtered frequency domain representation, wherein the reconstructed text image does not depict the layout elements; matching the reconstructed image to a form template; generating text from the reconstituted text image; and mapping a portion of the text to a corresponding field of the form template based on a position of the portion of the text in the reconstructed text image. ### Claim 5 (depends on 4) The system of claim 4, wherein the reconstructed image is matched to the form template using a neural network. ### Claim 6 (depends on 4) The system of claim 4, wherein the layout filter comprises a two dimensional high-pass filter. ### Claim 7 A non-transitory computer-readable medium comprising instructions which, when executed by one or more processors, performs an operation for removing text from form images, the operation comprising: generating a frequency domain representation of an image by applying a two dimensional frequency domain transformation to the image, wherein the image depicts at least a portion of a form and wherein the form includes layout elements and text elements; removing a portion of the frequency domain representation that represents the text elements by applying a first filter to the frequency domain representation, producing a first filtered frequency domain representation; generating a reconstructed image of at least the portion of the form by applying an inverse two dimensional frequency domain transformation to the first filtered frequency domain representation, wherein the reconstructed image does not depict the text elements; removing a portion of the frequency domain representation that represents the layout elements by applying a layout filter to the frequency domain representation, producing a second filtered frequency domain representation; generating a reconstructed text image by applying an inverse two dimensional frequency domain transformation to the second filtered frequency domain representation, wherein the reconstructed text image does not depict the layout elements; matching the reconstructed image to a form template; generating text from the reconstructed text image; and mapping a portion of the text to a corresponding field of the form template based on a position of the portion of the text in the reconstructed text image. ### Claim 8 (depends on 7) The non-transitory computer-readable medium of claim 7, wherein the operation further comprises: upon determining an amount of layout elements is below a threshold, adjusting the portion of the frequency domain removed by modifying the filter; and generating a modified reconstructed image by applying an inverse two dimensional frequency domain transformation to the frequency domain representation after being filtered with the modified filter. ### Claim 9 (depends on 8) The non-transitory computer-readable medium of claim 8, wherein the operation further comprises: generating a frequency domain representation of a second image by applying a two dimensional frequency domain transformation, wherein the second image depicts at least a second portion of the form; removing a portion of the frequency domain representation representing the text elements of the second image by applying a second filter to the frequency domain representation of the second image, producing a filtered frequency domain representation of the second image; and generating a reconstructed second image by applying an inverse two-dimensional frequency domain transformation to the filtered frequency domain representation of the second image, wherein the text element is not depicted in the reconstructed image; and generating a composite form image by combining the reconstructed image and the reconstructed second image. ### Claim 10 (depends on 7) The non-transitory computer-readable medium of claim 7, wherein the reconstructed image is matched to the form template using a neural network. ### Claim 11 (depends on 7) The non-transitory computer-readable medium of claim 7, wherein the layout filter comprises a two dimensional high-pass filter. ### Claim 12 (depends on 7) The non-transitory computer-readable medium of claim 7, wherein the two dimensional frequency domain transformation applied to the image is a two dimensional discrete Fourier transform.