Video quality evaluation method and apparatus, device, and storage medium

1. A video quality evaluation method, comprising:

acquiring an image sequence and audio information by decoding a to-be-evaluated video, wherein the to-be-evaluated video is a non-offending video;

extracting an action feature vector and a face feature vector from the image sequence, and extracting an audio feature vector from the audio information;

constructing a video feature vector according to at least one of the action feature vector, the face feature vector and the audio feature vector;

inputting the video feature vector into a first-layer evaluation network, and acquiring, by the first-layer evaluation network, a primary feature vector through recursive and non-linear operations;

inputting the primary feature vector into a second-layer evaluation network, and acquiring, by the second-layer evaluation network, an advanced feature vector through recursive and non-linear operations; and

acquiring a quality score of the to-be-evaluated video by performing, by an output layer, operations on the primary feature vector and the advanced feature vector,

wherein constructing the video feature vector according to at least one of the action feature vector, the face feature vector and the audio feature vector comprises:

setting the feature vector that fails to be extracted among the action feature vector, the face feature vector and the audio feature vector to zero; and

constructing the video feature vector by combining the action feature vector, the face feature vector and the audio feature vector.

2. The method according to claim 1 , wherein extracting the action feature vector from the image sequence comprises:

acquiring the action feature vector by extracting features from each frame of image by using an action feature extraction network;

and/or

extracting optical flow data of each frame of image, and

acquiring the action feature vector by extracting features from the optical flow data of each frame of image by using an action feature extraction network.

3. The method according to claim 2 , further comprising:

acquiring a training video set, and extracting image sequences in each training video;

acquiring an action classification network by training in a video database; and

acquiring the action feature extraction network by training the action classification network with the image sequences in each training video and deleting a classification layer from the action classification network.

4. The method according to claim 1 , wherein extracting the face feature vector from the image sequence comprises:

extracting the face feature vector from the image sequence by using a face feature extraction network.

5. The method according to claim 4 , wherein the face feature extraction network comprises a face detection sub-network and a feature extraction sub-network.

6. The method according to claim 5 , wherein extracting the face feature vector from the image sequence by using the face feature extraction network comprises:

detecting a face image in each frame of image in the image sequence by using the face detection sub-network; and

acquiring the face feature vector by extracting features the face image in each frame of image by using the feature extraction sub-network.

7. The method according to claim 6 , further comprising:

acquiring a training video set, and extracting face images in each training video;

acquiring a face recognition network by training in a face database; and

acquiring the feature extraction sub-network by training the face recognition network using the face images in each training video.

8. The method according to claim 1 , wherein extracting the audio feature vector from the audio information comprises:

extracting at least one of a cepstral coefficient and a spectrum of the audio information; and

acquiring the audio feature vector by extracting features from at least one of the cepstral coefficient and the spectrum by using an audio feature extraction network.

9. The method according to claim 8 , further comprising:

acquiring a training video set, and extracting the audio information in each training video;

acquiring an audio recognition network by training in an audio database; and

acquiring the audio feature extraction network by training the audio recognition network using the audio information in each training video.

10. A device, comprising:

at least one processor; and

a memory storing at least one program therein, wherein

the at least one program, when run by the at least one processor, causes the at least one processor to perform a video quality evaluation method, comprising:

acquiring an image sequence and audio information by decoding a to-be-evaluated video, wherein the to-be-evaluated video is a short video recorded by a user of a short video application;

extracting an action feature vector and a face feature vector from the image sequence, and extracting an audio feature vector from the audio information;

constructing a video feature vector according to at least one of the action feature vector, the face feature vector and the audio feature vector;

inputting the video feature vector into a first-layer evaluation network, and acquiring, by the first-layer evaluation network, a primary feature vector through recursive and non-linear operations;

inputting the primary feature vector into a second-layer evaluation network, and acquiring, by the second-layer evaluation network, an advanced feature vector through recursive and non-linear operations; and

acquiring a quality score of the to-be-evaluated video by performing, by an output layer, operations on the primary feature vector and the advanced feature vector,

wherein constructing the video feature vector according to at least one of the action feature vector, the face feature vector and the audio feature vector comprises:

setting the action feature vector, the face feature vector, or the audio feature vector to zero in response its failed extraction; and

constructing the video feature vector by combining the action feature vector, the face feature vector and the audio feature vector.

11. A non-transitory computer-readable storage medium storing a computer program, wherein the computer program, when run by a processing apparatus, causes the processing apparatus to perform a video quality evaluation method, comprising:

acquiring an image sequence and audio information by decoding a to-be-evaluated video, wherein the to-be-evaluated video is a non-offending video;

extracting an action feature vector and a face feature vector from the image sequence, and extracting an audio feature vector from the audio information;

constructing a video feature vector according to at least one of the action feature vector, the face feature vector and the audio feature vector;

inputting the video feature vector into a first-layer evaluation network, and acquiring, by the first-layer evaluation network, a primary feature vector through recursive and non-linear operations;

inputting the primary feature vector into a second-layer evaluation network, and acquiring, by the second-layer evaluation network, an advanced feature vector through recursive and non-linear operations; and

acquiring a quality score of the to-be-evaluated video by performing, by an output layer, operations on the primary feature vector and the advanced feature vector,

wherein constructing the video feature vector according to at least one of the action feature vector, the face feature vector and the audio feature vector comprises:

setting the feature vector that fails to be extracted among the action feature vector, the face feature vector and the audio feature vector to zero; and

constructing the video feature vector by combining the action feature vector, the face feature vector and the audio feature vector.

12. The device according to claim 10 , wherein extracting the action feature vector from the image sequence comprises:

acquiring the action feature vector by extracting features from each frame of image by using an action feature extraction network;

and/or

extracting optical flow data of each frame of image; and

acquiring the action feature vector by extracting features from the optical flow data of each frame of image by using an action feature extraction network.

13. The device according to claim 12 , wherein the method further comprises:

acquiring a training video set, and extracting image sequences in each training video;

acquiring an action classification network by training in a video database; and

acquiring the action feature extraction network by training the action classification network with the image sequences in each training video and deleting a classification layer from the action classification network.

14. The device according to claim 10 , wherein extracting the face feature vector from the image sequence comprises:

extracting the face feature vector from the image sequence by using a face feature extraction network.

15. The device according to claim 14 , wherein the face feature extraction network comprises a face detection sub-network and a feature extraction sub-network.

16. The device according to claim 15 , wherein extracting the face feature vector from the image sequence by using the face feature extraction network comprises:

detecting a face image in each frame of image in the image sequence by using the face detection sub-network; and

acquiring the face feature vector by extracting features the face image in each frame of image by using the feature extraction sub-network.

17. The device according to claim 16 , wherein the method further comprises:

acquiring a training video set, and extracting face images in each training video;

acquiring a face recognition network by training in a face database; and

acquiring the feature extraction sub-network by training the face recognition network using the face images in each training video.

18. The device according to claim 10 , wherein extracting the audio feature vector from the audio information comprises:

extracting at least one of a cepstral coefficient and a spectrum of the audio information; and

acquiring the audio feature vector by extracting features from at least one of the cepstral coefficient and the spectrum by using an audio feature extraction network.

19. The device according to claim 18 , wherein the method further comprises:

acquiring a training video set, and extracting the audio information in each training video;

acquiring an audio recognition network by training in an audio database; and

acquiring the audio feature extraction network by training the audio recognition network using the audio information in each training video.