Apparatus and method for detecting scene text in an image

1. A method comprising:

identifying, by one or more processors, text regions in an image via an automatically learned attention map and by conducting pixel-wise review of text;

aggregating, by the one or more processors, multi-scale inception features;

generating, by the one or more processors and based at least in part on the multi-scale inception features, a set of aggregated inception features; and

generating, by the one or more processors and using at least the set of aggregated inception features, word-level text bounding boxes in the image.

2. The method of claim 1 , wherein conducting the pixel-wise review of text comprises:

convoluting, by the one or more processors, the image as a whole with a convolutional neural network to generate intermediate features;

generating, by the one or more processors, the attention map via a softmax function that uses the intermediate features; and

performing, by the one or more processors, pixel-wise multiplication between the attention map and the convolutional features to suppress background interference.

3. The method of claim 2 , wherein performing the pixel-wise multiplication comprises masking the background either by features or raw images.

4. The method of claim 1 , further comprising:

aggregating, by the one or more processors, convolutional features;

capturing, by the one or more processors, context information by using multi-scale receptive fields; and

aggregating, by the one or more processors, multi-layer inception modules and enhancing the convolutional features towards text task.

5. The method of claim 1 , wherein (a) identifying the text regions is performed by a text attention module, and (b) aggregating multi-scale inception features is performed by a hierarchical inception module.

6. The method of claim 5 , wherein (a) generating the set of aggregated inception features is performed by the hierarchical inception module and (b) generating the word-level text bounding boxes in the image is performed by a word prediction module.

7. A apparatus comprising at least a processor, and a memory associated with the processor having computer coded instructions therein, with the computer instructions configured to, when executed by the processor, cause the apparatus to:

identify text regions in an image via an automatically learned attention map and by conducting pixel-wise review of text;

aggregate multi-scale inception features;

generate, based at least in part on the multi-scale inception features, a set of aggregated inception features; and

generate, using at least the set of aggregated inception features, word-level text bounding boxes in the image.

8. The apparatus of claim 7 , wherein the conducting the pixel-wise review of text comprises:

convolute the image as a whole with a convolutional neural network to generate intermediate features;

generate the attention map via a softmax function that uses the intermediate features; and

perform pixel-wise multiplication between the attention map and the convolutional features to suppress background interference.

9. The apparatus of claim 8 , wherein performing the pixel-wise multiplication comprises masking the background either by features or raw images.

10. The apparatus of claim 7 , further comprising:

aggregate convolutional features;

capture context information by using multi-scale receptive fields; and

aggregate multi-layer inception modules and enhancing the convolutional features towards text task.

11. The apparatus of claim 7 , wherein (a) identifying the text regions is performed by a text attention module, and (b) aggregating multi-scale inception features is performed by a hierarchical inception module.

12. The apparatus of claim 7 , wherein (a) generating the set of aggregated inception features is performed by the hierarchical inception module and (b) generating the word-level text bounding boxes in the image is performed by a word prediction module.

13. A computer program product comprising at least one non-transitory computer-readable storage medium having computer-readable program code portions stored therein, the computer-readable program code portions comprising:

an executable portion configured to identify text regions in an image via an automatically learned attention map and by conducting pixel-wise review of text;

an executable portion configured to aggregate multi-scale inception features;

an executable portion configured to generate, based at least in part on the multi-scale inception features, a set of aggregated inception features; and

an executable portion configured to generate, using at least the set of aggregated inception features, word-level text bounding boxes in the image.

14. The computer program product of claim 13 , wherein the conducting the pixel-wise review of text comprises:

convolute the image as a whole with a convolutional neural network to generate intermediate features;

generate the attention map via a softmax function that uses the intermediate features; and

perform pixel-wise multiplication between the attention map and the convolutional features to suppress background interference.

15. The computer program product of claim 14 , wherein performing the pixel-wise multiplication comprises masking the background either by features or raw images.

16. The computer program product of claim 13 , further comprising:

an executable portion aggregate convolutional features;

an executable portion capture context information by using multi-scale receptive fields; and

an executable portion aggregate multi-layer inception modules and enhancing the convolutional features towards text task.

17. The computer program product of claim 13 , wherein (a) identifying the text regions is performed by a text attention module, and (b) aggregating multi-scale inception features is performed by a hierarchical inception module.

18. The computer program product of claim 13 , wherein (a) generating the set of aggregated inception features is performed by the hierarchical inception module and (b) generating the word-level text bounding boxes in the image is performed by a word prediction module.