Image description generation method, apparatus and system, and medium and electronic device

1. An image description generation method, comprising:

acquiring one or more image region features in a target image, and obtaining a current input vector by performing a mean pooling on the image region features;

obtaining respective outer product vectors of the image region features by respectively linearly fusing the current input vector and each of the image region features;

calculating, based on the respective outer product vectors of the image region features, an attention distribution of the image region features in a spatial dimension and an attention distribution of the image region features in a channel dimension; and

generating an image description of the target image based on the attention distribution of the image region features in the spatial dimension and the attention distribution of the image region features in the channel dimension.

2. The method according to claim 1 , wherein the acquiring image region features in the target image comprises:

obtaining at least one object included in the target image by detecting the target image; and

obtaining the image region features corresponding to each object included in the target image.

3. The method according to claim 1 , wherein the calculating, based on the respective outer product vectors of the image region features, the attention distribution of the image region features in the spatial dimension and the attention distribution of the image region features in the channel dimension comprises:

obtaining the attention distribution of the image region features in the spatial dimension by mapping the respective outer product vectors of the image region features; and

obtaining the attention distribution of the image region features in the channel dimension by performing a squeeze-excitation processing on the respective outer product vectors of the image region features.

4. The method according to claim 3 , wherein the obtaining the attention distribution of the image region features in the spatial dimension by mapping the respective outer product vectors of the image region features comprises:

obtaining an attention weight corresponding to each image region by mapping each of the outer product vectors of the image region features; and

obtaining the attention distribution of the image region features in the spatial dimension by normalizing the attention weight corresponding to each image region.

5. The method according to claim 3 , wherein the obtaining the attention distribution of the image region features in the channel dimension by performing the squeeze-excitation processing on the respective outer product vectors of the image region features comprises:

obtaining a squeezed channel expression by fusing each of the outer product vectors of the image region features in the spatial dimension;

obtaining the attention weight of the image region features in the channel dimension by performing an excitation processing on the squeezed channel expression; and

obtaining the attention distribution of the image region features in the channel dimension by performing a regularization processing on the attention weight of the image region features in each channel dimension.

6. The method according to claim 1 , wherein the generating an image description of the target image based on the attention distribution in the spatial dimension and the attention distribution in the channel dimension comprises:

obtaining respective current aggregated features by fusing, based on the attention distribution of the image region features in the spatial dimension, the image region features and the respective outer product vectors of the image region features;

obtaining a target aggregated feature by fusing the respective current aggregated features based on the attention distribution of the image region features in the channel dimension; and

generating the image description of the target image based on the target aggregated feature.

7. The method according to claim 1 , wherein the obtaining respective outer product vectors of the image region features by linearly fusing the current input vector and the image region features comprises:

obtaining the respective outer product vectors of the image region features by respectively performing an exponential mapping on the current input vector and each of the image region features, and performing a linear fusion on the exponential mapped current input vector and the exponential mapped image region feature.

8. The method according to claim 1 , wherein the respective outer product vectors of the image region features comprises second-order feature interactive information between a current hidden state feature of a decoder and an encoded image region feature.

9. An image description generation system implementing the image description generation method according to claim 1 , comprising:

an object detection module configured to acquire one or more image region features in a target image;

an encoder connected to the object detection module via a network, and comprising a plurality of high-order attention modules;

a decoder connected with the encoder via the network and comprising a plurality of long and short-term memory networks with a high-order attention mechanism;

wherein the plurality of high-level attention modules are configured to:

acquire the one or more image region features in the target image, and obtain a current input vector by performing a mean pooling on the image region features;

obtain respective outer product vectors of the image region features by respectively linearly fusing the current input vector and each of the image region features;

calculate, based on the respective outer product vectors of the image region features, an attention distribution of the image region features in a spatial dimension and an attention distribution of the image region features in a channel dimension; and

generate an image description of the target image based on the attention distribution of the image region features in the spatial dimension and the attention distribution of the image region features in the channel dimension.

10. A non-transitory computer-readable storage medium on which a computer program is stored, wherein when the computer program is executed by a processor, an image description generation method is implemented,

wherein the image description generation method comprises:

acquiring one or more image region features in a target image, and obtaining a current input vector by performing a mean pooling on the image region features;

obtaining respective outer product vectors of the image region features by respectively linearly fusing the current input vector and each of the image region features;

calculating, based on the respective outer product vectors of the image region features, an attention distribution of the image region features in a spatial dimension and an attention distribution of the image region features in a channel dimension; and

generating an image description of the target image based on the attention distribution of the image region features in the spatial dimension and the attention distribution of the image region features in the channel dimension.

11. An electronic device, comprising:

a processor; and

memory for storing instructions executed by the processor;

wherein, the processor is configured to:

acquire one or more image region features in a target image, and obtain a current input vector by performing a mean pooling on the image region features;

obtain respective outer product vectors of the image region features by respectively linearly fusing the current input vector and each of the image region features;

calculate, based on the respective outer product vectors of the image region features, an attention distribution of the image region features in a spatial dimension and an attention distribution of the image region features in a channel dimension; and

generate an image description of the target image based on the attention distribution of the image region features in the spatial dimension and the attention distribution of the image region features in the channel dimension.

12. The electronic device according to claim 11 , wherein the processor is configured to:

obtain at least one object included in the target image by detecting the target image; and

obtain the image region features corresponding to each object included in the target image.

13. The electronic device according to claim 11 , wherein the processor is configured to:

obtain the attention distribution of the image region features in the spatial dimension by mapping the respective outer product vectors of the image region features; and

obtain the attention distribution of the image region features in the channel dimension by performing a squeeze-excitation processing on the respective outer product vectors of the image region features.

14. The electronic device according to claim 13 , wherein the processor is configured to:

obtain an attention weight corresponding to each image region by mapping each of the outer product vectors of the image region features; and

obtain the attention distribution of the image region features in the spatial dimension by normalizing the attention weight corresponding to each image region.

15. The electronic device according to claim 13 , wherein the processor is configured to:

obtain a squeezed channel expression by fusing each of the outer product vectors of the image region features in the spatial dimension;

obtain the attention weight of the image region features in the channel dimension by performing an excitation processing on the squeezed channel expression; and

obtain the attention distribution of the image region features in the channel dimension by performing a regularization processing on the attention weight of the image region features in each channel dimension.

16. The electronic device according to claim 11 , wherein the processor is configured to:

obtain respective current aggregated features by fusing, based on the attention distribution of the image region features in the spatial dimension, the image region features and the respective outer product vectors of the image region features;

obtain a target aggregated feature by fusing the respective current aggregated features based on the attention distribution of the image region features in the channel dimension; and

generate the image description of the target image based on the target aggregated feature.

17. The electronic device according to claim 11 , wherein the processor is configured to:

obtain the respective outer product vectors of the image region features by respectively performing an exponential mapping on the current input vector and each of the image region features, and performing a linear fusion on the exponential mapped current input vector and the exponential mapped image region feature.

18. The electronic device according to claim 11 , wherein the respective outer product vectors of the image region features comprises second-order feature interactive information between a current hidden state feature of a decoder and an encoded image region feature.