Automatic determination of the presence of burn-in overlay in video imagery
Systems, methods and computer systems for the automatic determination of presence or absence of burn-in overlay data are provided. The systems, methods, and computer systems implement mask generation, edge detection, feature vector generation methods that are combined with machine learning classifiers to rapidly and automatically determine the presence or absence of burn-in overlays in the image for the purpose of removal or other forms to obfuscate burn-in overlay data so as to maintain confidential or classified information while allowing for the release of remaining image data.
1 . A computing system for identifying the presence of burn-in in video, the computing system comprising:
a processor; and
a memory coupled to the processor, the memory storing instructions which, when executed by the processor, cause the computing system to perform operations comprising:
receiving an image;
applying a masking filter to the image to generate a masked image;
calculating, by an edge detection filter applied to the masked image, the presence of a plurality of vertical and horizontal features in the masked image for a plurality of pixels in the masked image;
generating a feature vector for the plurality of pixels in the masked image; wherein the feature vector is indicative of the presence and position of vertical and horizontal features;
predicting, by a machine learning algorithm, whether the feature vector is indicative of the presence of a burn-in; and
initiating a redaction process to redact pixels corresponding to the feature vector when the machine learning algorithm predicts that the feature vector is indicative of the presence of burn-in.
2 . The computing system of claim 1 , wherein the edge detection filter is a Haar filter.
3 . The computing system of claim 1 , wherein the machine learning algorithm is a Bayesian machine learning algorithm.
4 . The computing system of claim 3 , wherein the Bayesian machine learning algorithm is a Relevance Vector Machine.
5 . The computing system of claim 1 , wherein the machine learning algorithm is a binary classifier.
6 . The computing system of claim 1 , wherein the images are received at a rate of at least 30 images per second or less.
7 . The computing system of claim 1 , wherein the redaction process further comprises the operation of replacing pixels corresponding to the mask in the image that has been predicted to include burn-in.
8 . The computing system of claim 2 , wherein the Haar filter performs the operations of:
calculating vertical and horizontal Haar features for the plurality of pixels in the masked image by computing vertical and horizontal Haar integral kernel responses on the integral image of the mask; calculating a summed Haar filter response by summing the vertical and horizontal Haar features;
thresholding pixels from the plurality of pixels based on whether the summed Haar response exceeds a predetermined threshold and
resizing and reshaping the result of the thresholding into a feature vector.
9 . At least one non-transitory computer readable medium comprising instructions which, when executed by a computing system, cause the computing system to perform operations comprising:
applying a filter to an image to generate a masked image;
calculating, by an edge detection filter applied to the masked image, the presence of a plurality of vertical and horizontal features in the masked image for the plurality of pixels in the masked image;
generating a feature vector for the plurality of pixels in the masked image; that describes the presence and position of vertical and horizontal features;
predicting, by a machine learning algorithm, whether the feature vector is indicative of the presence of a burn-in; and
performing a redaction process to redact pixels corresponding to the feature vector when the machine learning algorithm predicts that the feature vector is indicative of the presence of burn-in.
10 . The at least one non-transitory computer readable medium of claim 9 ,
wherein the edge detection filter applied to the masked edge is a Haar filter.
11 . The at least one non-transitory computer readable medium of claim 9 ,
wherein the machine learning algorithm is a Bayesian machine learning algorithm.
12 . The at least one non-transitory computer readable medium of claim 11 ,
wherein the Bayesian machine learning algorithm is a Relevance Vector Machine.
13 . The at least one non-transitory computer readable medium of claim 9 ,
wherein the machine learning algorithm is a binary classifier.
14 . The at least one non-transitory computer readable medium of claim 9 ,
wherein the redaction process further comprises replacing pixels corresponding to the mask in the image indicating the presence of burn-in.
15 . The at least one non-transitory computer readable medium of claim 10 ,
wherein the Haar filter calculates vertical and horizontal Haar features for the plurality of pixels in the masked image by computing vertical and horizontal Haar integral kernel responses on the integral image of the mask;
calculates a summed Haar filter response by summing the vertical and horizontal Haar features; and
thresholds pixels from the plurality of pixels based on whether the summed Haar response exceeds a predetermined threshold and resizes and reshapes the result of the thresholding into a feature vector.
16 . A method comprising:
applying a filter to an image to generate a masked image;
calculating, by an edge detection filter applied to the masked image, the presence of a plurality of vertical and horizontal features in the masked image for the plurality of pixels in the masked image;
generating a feature vector for the plurality of pixels in the masked image that describes the presence and position of vertical and horizontal features; predicting, by a machine learning algorithm, whether the feature vector is indicative of the presence of a burn-in; and
initiating a redaction process to redact pixels corresponding to the feature vector when the machine learning algorithm predicts that the feature vector is indicative of the presence of burn-in.
17 . The method of claim 16 wherein the edge detection filter is a Haar filter.
18 . The method of claim 16 wherein the machine learning algorithm is a Bayesian machine learning algorithm.
19 . The method of claim 18 wherein the machine learning algorithm is one of i) a Relevance Vector Machine; or ii) a binary classifier.
20 . The method of claim 16 wherein the edge detection filter is a Haar filter and wherein the Haar filter calculates vertical and horizontal Haar features for the plurality of pixels in the masked image by computing vertical and horizontal Haar integral kernel responses on the integral image of the mask;
calculates a summed Haar response by summing the vertical and horizontal Haar features; and
thresholds pixels from the plurality of pixels based on whether the summed Haar response exceeds a predetermined threshold and
resizes and reshapes the result of the thresholding into a feature vector.