Recurrent deep neural network system for detecting overlays in images
View Patent ↗In one aspect, an example method includes a processor (1) applying a feature map network to an image to create a feature map comprising a grid of vectors characterizing at least one feature in the image and (2) applying a probability map network to the feature map to create a probability map assigning a probability to the at least one feature in the image, where the assigned probability corresponds to a likelihood that the at least one feature is an overlay. The method further includes the processor determining that the probability exceeds a threshold, and responsive to the processor determining that the probability exceeds the threshold, performing a processing action associated with the at least one feature.
1. A method comprising:
a processor applying a feature map network to an image to create a feature map, wherein the feature map comprises a grid of vectors characterizing at least one feature in the image;
the processor applying a probability map network to the feature map to create a probability map assigning a probability to the at least one feature in the image, wherein the assigned probability corresponds to a likelihood that the at least one feature is an overlay;
the processor determining that the assigned probability exceeds a threshold probability value; and
responsive to the processor determining that the assigned probability exceeds the threshold probability value, the processor (i) determining compliance with an advertising requirement, (ii) performing optical character recognition (OCR) on the at least one feature, (iii) determining a channel associated with the at least one feature, (iv) determining a show associated with the at least one feature, (v) determining a hyperlink associated with the at least one feature, or (vi) accessing a hyperlinked object associated with the at least one feature.
2. The method of claim 1 , further comprising:
responsive to the processor determining that the assigned probability exceeds the threshold probability value, the processor (i) removing the at least one feature from the image or (ii) inpainting the at least one feature.
3. The method of claim 1 , further comprising:
training the feature map network prior to the processor applying the feature map network to the image.
4. The method of claim 3 , wherein training the feature map network comprises generating a plurality of groundtruth masks from a corresponding plurality of training images, wherein each training image of the plurality of training images has a respective overlay.
5. The method of claim 4 , further comprising:
creating the plurality of training images by placing a plurality of overlay images onto an original image set, wherein each overlay image of the plurality of overlay images is semi-transparent or opaque.
6. The method of claim 5 , further comprising:
varying the plurality of overlay images by at least one of grayscale intensity, color hue, color value, color intensity, opacity, size, or position in the original image set.
7. The method of claim 1 , wherein applying the feature map network to the image comprises applying a plurality of convolutional layers and max-pooling layers to the image.
8. The method of claim 7 , wherein the image is described on a per-pixel basis using RGB values, and wherein the plurality of convolutional layers and max-pooling layers comprises the following, in order:
two convolutional 3×3 64-length layers;
a first max-pooling 2×2 layer;
two convolution 3×3 128-length layers;
a second max-pooling 2×2 layer;
three convolution 3×3 256-length layers;
a third max-pooling 2×2 layer;
three convolution 3×3 512-length layers;
a fourth max-pooling 2×2 layer; and
two convolution 1×1 4096-length layers.
9. The method of claim 1 , wherein applying the probability map network to the feature map comprises applying the following, in order:
a first LSTM across rows and a second LSTM across columns;
two fully connected layers;
two sigmoid layers; and
a multiplication layer.
10. A non-transitory computer-readable medium having stored thereon program instructions that, upon execution by a processor, cause the processor to perform a set of acts comprising:
applying a feature map network to an image to create a feature map, wherein the feature map includes data characterizing a feature in the image relative to other features in the image;
applying a probability map network to the feature map to create a probability map assigning a probability to the feature in the image, wherein the assigned probability corresponds to a likelihood that the feature is an overlay;
determining that the assigned probability exceeds a threshold probability value; and
responsive to determining that the assigned probability exceeds the threshold probability value, (i) determining compliance with an advertising requirement, (ii) performing optical character recognition (OCR) on the at least one feature, (iii) determining a channel associated with the at least one feature, (iv) determining a show associated with the at least one feature, (v) determining a hyperlink associated with the at least one feature, or (vi) accessing a hyperlinked object associated with the at least one feature.
11. The non-transitory computer-readable medium of claim 10 , the set of acts further comprising:
responsive to determining that the assigned probability exceeds the threshold probability value, (i) removing the at least one feature from the image or (ii) inpainting the at least one feature.
12. The non-transitory computer-readable medium of claim 10 , the set of acts further comprising:
training the feature map network prior to the processor applying the feature map network to the image, wherein training the feature map network comprises generating a plurality of groundtruth masks from a corresponding plurality of training images, wherein each training image of the plurality of training images has a respective overlay.
13. The non-transitory computer-readable medium of claim 12 , wherein each groundtruth mask of the plurality of groundtruth masks is the same size as its corresponding training image, the set of acts further comprising:
setting each pixel in a respective groundtruth mask of the plurality of groundtruth masks to 1 if the pixel is overlapped by the respective overlay or to 0 if the pixel is not overlapped by the respective overlay.
14. The non-transitory computer-readable medium of claim 12 , the set of acts further comprising:
creating the plurality of training images by placing a plurality of overlay images onto an original image set, wherein each overlay image of the plurality of overlay images is semi-transparent or opaque.
15. The non-transitory computer-readable medium of claim 10 , wherein applying the feature map network to the image comprises applying a plurality of convolutional layers and max-pooling layers to the image.
16. The non-transitory computer-readable medium of claim 15 , wherein the image is described on a per-pixel basis using RGB values, and wherein the plurality of convolutional layers and max-pooling layers comprises the following, in order:
two convolutional 3×3 64-length layers;
a first max-pooling 2×2 layer;
two convolution 3×3 128-length layers;
a second max-pooling 2×2 layer;
three convolution 3×3 256-length layers;
a third max-pooling 2×2 layer;
three convolution 3×3 512-length layers;
a fourth max-pooling 2×2 layer; and
two convolution 1×1 4096-length layers.
17. The non-transitory computer-readable medium of claim 10 , wherein applying the probability map network to the feature map comprises applying bi-directional LSTMs with peephole connections to the feature map.
18. The non-transitory computer-readable medium of claim 17 , wherein the bi-directional LSTMs with peephole connections comprises the following, in order:
a first LSTM across rows and a second LSTM across columns;
two fully connected layers;
two sigmoid layers; and
a multiplication layer.
19. A system, comprising:
a processor; and
a non-transitory computer-readable medium having stored thereon program instructions that, upon execution by the processor, cause the processor to perform a set of acts comprising:
training a feature map network by generating a plurality of groundtruth masks from a corresponding plurality of training images, wherein each of the plurality of training images has an overlay with at least one of a respective grayscale intensity, color hue, color value, color intensity, opacity, size, or position that differs relative to other overlays in the plurality of training images;
applying the trained feature map network to an image to create a feature map, wherein the feature map includes data characterizing a feature in the image relative to other features in the image, wherein the feature map network comprises a very deep convolutional neural network having a plurality of convolutional layers and max-pooling layers;
applying a probability map network to the feature map to create a probability map, wherein the probability map network comprises at least two recurrent layers including at least one vertical recurrent layer and at least one horizontal recurrent layer that are multiplied to provide an output indicative of a probability that the feature in the image is an overlay;
determining that the probability exceeds a threshold probability value; and
responsive to determining that the probability exceeds the threshold probability value, (i) determining compliance with an advertising requirement, (ii) performing optical character recognition (OCR) on the feature, (iii) determining a channel associated with the feature, (iv) determining a show associated with the feature, (v) determining a hyperlink associated with the feature, or (vi) accessing a hyperlinked object associated with the feature.
20. The system of claim 19 , the set of acts further comprising:
responsive to determining that the probability exceeds the threshold probability value, (i) removing the feature from the image or (ii) inpainting the feature.