Methods and apparatus for end-to-end unsupervised multi-document blind image denoising
Methods and apparatus for end-to-end unsupervised multi-document blind image denoising is presented. The multi-document blind image denoiser removes various noise types from noisy documents without paired target cleaned documents and preserves the contents for optical character recognition. The end-to-end unsupervised multi-document blind image denoiser integrates a Mixture of Experts with a cycle-consistent GAN as the base network that effectively removes multiple types of noise, including salt & pepper noise, blurred and/or faded text, as well as watermarks from documents at various levels of intensity.
1 . An apparatus comprising:
a processor; and
a memory operatively coupled to the processor, the memory storing instructions to cause the processor to:
receive a noisy image without a paired clean image, the noisy image being associated with a document type;
input the noisy image to a first machine learning model layer to generate a blind denoised image;
input the blind denoised image to a second machine learning model layer to generate a first classification of the document type;
input the blind denoised image to a third machine learning model layer to generate a blind noised image;
input the blind noised image to a fourth machine learning model layer to generate a second classification of the document type; and
generate a clean image based on a comparison between the first classification of the document type and the second classification of the document type.
2 . The apparatus of claim 1 , wherein the memory stores instructions to further cause the processor to:
classify the noisy image based on an image structure type; and
store the noisy image in a database categorized by the image structure type.
3 . The apparatus of claim 1 , wherein the document type includes one of a structured document type, a semi-structured document type, or an unsupervised document type.
4 . The apparatus of claim 1 , wherein the memory stores instructions to further cause the processor to:
input the blind denoised image to a fifth machine learning model layer to determine a first noise intensity value; and
input the blind noised image to a sixth machine learning model layer to determine a second noise intensity value, the clean image being generated based further on a comparison between the first noise intensity value and the second noise intensity value.
5 . The apparatus of claim 1 , wherein the memory stores instructions to further cause the processor to:
input the blind denoised image to a fifth machine learning model layer to generate a first classification of a noise type; and
input the blind noised image to a sixth machine learning model layer to generate a second classification of the noise type, the clean image being generated based further on a comparison between the first classification of the noise type and the second classification of the noise type.
6 . The apparatus of claim 1 , wherein the memory stores instructions to further cause the processor to:
input the blind denoised image to a fifth machine learning model to classify the blind denoised image as having at least one of salt & pepper (S&P) noise, blurred text, faded text, a shading, or a watermark; and
input the blind noised image to a sixth machine learning model layer to classify the blind noised image as having the at least one of the salt & pepper (S&P) noise, the blurred text, the faded text, the shading, or the watermark, the clean image being generated based further on the classifying the blind denoised image and the classifying the blind noised image.
7 . The apparatus of claim 1 , further including:
a plurality of denoising gates, each denoising gate of the plurality of denoising gates configured to activate at least one denoising channel of a plurality of denoising channels of the first machine learning model layer;
a plurality of noising gates, each noising gate of the plurality of noising gates configured to activate at least one noising channel of a plurality of noising channels of the third machine learning model layer; and
an embedder, the memory storing instructions to cause the processor to:
compute, via the embedder, a denoising mixture weight of a plurality of denoising mixture weights for each denoising gate of the plurality of denoising gates based on the noisy image; and
compute, via the embedder, a noising mixture weight of a plurality of noising mixture weights for each noising gate of the plurality of noising gates based on the blind denoised image.
8 . The apparatus of claim 1 , wherein:
the clean image is a clean instance of the noisy image.
9 . The apparatus of claim 1 , wherein:
an end-to-end machine learning model includes the first machine learning model layer, the second machine learning model layer, the third machine learning model layer, and the fourth machine learning model layer.
10 . A method, comprising:
receiving, at a processor, a noisy image without a paired clean image;
inputting, via the processor, the noisy image to a first machine learning model layer to generate a blind denoised image;
inputting, via the processor, the blind denoised image to a second machine learning model layer to generate a first classification of at least one of an image structure type, a noise type, or a denoising intensity;
inputting, via the processor, the blind denoised image to a third machine learning model layer to generate a blind noised image;
inputting, via the processor, the blind noised image to a fourth machine learning model layer to generate a second classification of the at least one of the image structure type, the noise type, or the denoising intensity; and
generating, via the processor, a clean image based on a comparison between the first classification and the second classification.
11 . The method of claim 10 , wherein the image structure type includes one of a structured document type, a semi-structured document type, or an unsupervised document type.
12 . The method of claim 10 , further comprising:
inputting, via the processor, the blind denoised image to a fifth machine learning model layer to determine a first noise intensity value; and
inputting, via the processor, the blind noised image to a sixth machine learning model layer to determine a second noise intensity value, the clean image being generated based further on a comparison between the first noise intensity value and the second noise intensity value.
13 . The method of claim 10 , further comprising:
inputting, via the processor, the blind denoised image to a fifth machine learning model layer to generate a first classification of a noise type; and
inputting, via the processor, the blind noised image to a sixth machine learning model layer to generate a second classification of the noise type, the clean image being generated based further on a comparison between the first classification of the noise type and the second classification of the noise type.
14 . The method of claim 10 , further comprising:
inputting, via the processor, the blind denoised image to a fifth machine learning model to classify the blind denoised image as having at least one of salt & pepper (S&P) noise, blurred text, faded text, a shading, or a watermark; and
inputting, via the processor, the blind noised image to a sixth machine learning model layer to classify the blind noised image as having the at least one of the salt & pepper (S&P) noise, the blurred text, the faded text, the shading, or the watermark, the clean image being generated based further on the classifying the blind denoised image and the classifying the blind noised image.
15 . The method of claim 10 , wherein:
the inputting the noisy image to the first machine learning model layer includes:
computing, via the processor, a denoising mixture weight for a denoising gate of the first machine learning model layer based on the noisy image, and
activating, via the processor, a denoising channel using the denoising gate that has the denoising mixture weight, to generate the blind denoised image; and
the inputting the blind denoised image to the third machine learning model layer includes:
computing, via the processor, a noising mixture weight for a noising gate of the third machine learning model layer based on the blind denoised image, and activating, via the processor, a noising channel using the noising gate that has the noising mixture weight, to generate the blind noised image.
16 . A non-transitory, processor-readable medium storing instructions that, when executed by a processor, cause the processor to:
receive a noisy image without a paired clean image;
input the noisy image to a first machine learning model layer to generate a blind denoised image;
input the blind denoised image to a second machine learning model layer to generate a first classification of at least one of an image structure type, a noise type, or a denoising intensity;
input the blind denoised image to a third machine learning model layer to generate a blind noised image;
input the blind noised image to a fourth machine learning model layer to generate a second classification of the at least one of the image structure type, the noise type, or the denoising intensity; and
generate a clean image based on a comparison between the first classification and the second classification.
17 . The non-transitory, processor-readable medium of claim 16 , wherein the image structure type includes one of a structured document type, a semi-structured document type, or an unsupervised document type.
18 . The non-transitory, processor-readable medium of claim 16 , further storing instructions to cause the processor to:
input the blind denoised image to a fifth machine learning model layer to determine a first noise intensity value; and
input the blind noised image to a sixth machine learning model layer to determine a second noise intensity value, the clean image being generated based further on a comparison between the first noise intensity value and the second noise intensity value.
19 . The non-transitory, processor-readable medium of claim 16 , further storing instructions to cause the processor to:
input the blind denoised image to a fifth machine learning model layer to generate a first classification of a noise type; and
input the blind noised image to a sixth machine learning model layer to generate a second classification of the noise type, the clean image being generated based further on a comparison between the first classification of the noise type and the second classification of the noise type.
20 . The non-transitory, processor-readable medium of claim 16 , further storing instructions to cause the processor to:
input the blind denoised image to a fifth machine learning model to classify the blind denoised image as having at least one of salt & pepper (S&P) noise, blurred text, faded text, a shading, or a watermark; and
input the blind noised image to a sixth machine learning model layer to classify the blind noised image as having the at least one of the salt & pepper (S&P) noise, the blurred text, the faded text, the shading, or the watermark, the clean image being generated in response to classifying the blind denoised image and classifying the blind noised image.