Method, device and computer readable medium for intrinsic popularity evaluation and content compression based thereon

1. A method of content compression, the method comprising:

compressing information with a variable compression rate corresponding to an intrinsic popularity score of a content of the information, wherein the content of the information comprises an image; and

before the information is compressed, determining the intrinsic popularity score of the image using a deep neural network (DNN) based intrinsic popularity assessment model;

wherein the compressing of the information comprises:

classifying the intrinsic popularity score into a popularity level; and

compressing the image with the variable compression rate, the variable compression rate corresponding to the popularity level, the variable compression rate being in a negative correlation to the corresponding intrinsic popularity score, wherein the image is compressed by using a DNN based autoencoder, the DNN based autoencoder comprising learnable weights, the learnable weights being learnt according to an objective function given by = r +λ d , where r is a coding cost, d is a reconstruction error, and λ is a Lagrange multiplier for controlling rate-distortion trade-off, and wherein A is a monotonic increasing function of the popularity level such that more bits are allocated for more popular images to thereby enable the image to be compressed with the variable compression rate corresponding to the popularity level and in the negative correlation to the corresponding intrinsic popularity score.

2. The method according to claim 1 , further comprising:

training the DNN based intrinsic popularity assessment model, wherein the training comprises:

retrieving a plurality of historical images;

constructing a plurality of popularity-discriminable image pairs (PDIPs) based on the plurality of historical images, wherein each PDIP comprises a first image and a second image, the first image having a probability of higher intrinsic popularity than the second image; and

training the DNN based intrinsic popularity assessment model with the plurality of PDIPs.

3. The method according to claim 2 , wherein the training of the DNN based intrinsic popularity assessment model with the plurality of PDIPs comprises:

for each PDIP of the plurality of PDIPs,

determining a first intrinsic popularity score for the first image and a second intrinsic popularity score for the second image using the DNN based intrinsic popularity assessment model, and

optimising the DNN based intrinsic popularity assessment model by minimizing a binary cross entropy loss between a score difference between the first intrinsic popularity score and the second intrinsic popularity score and a ground truth binary label denoting whether the first image is intrinsically more popular than the second image.

4. The method according to claim 1 , wherein the intrinsic popularity score for the image is associated with a visual content of the image.

5. The method according to claim 4 , wherein the intrinsic popularity score for the image is further associated with one or more attributes of the image, the one or more attributes including:

upload/download time,

post URL,

user ID,

type of the visual content,

caption content, and/or

number of comments.

6. A device for content compression, the device comprising:

at least one processor configurable to allocate resources; and

a memory including computer program code for execution by the at least one processor, wherein the computer program code instructs the at least one processor to allocate resources by:

compressing information with a variable compression rate corresponding to an intrinsic popularity score of a content of the information, wherein the content of the information comprises an image; and

before the information is compressed, determining the intrinsic popularity score of the image using a deep neural network (DNN) based intrinsic popularity assessment model;

wherein the compressing of the information comprises:

classifying the intrinsic popularity score into a popularity level; and

compressing the image with the variable compression rate, the variable compression rate corresponding to the popularity level, the variable compression rate being in a negative correlation to the corresponding intrinsic popularity score, wherein the image is compressed by using a DNN based autoencoder, the DNN based autoencoder comprising learnable weights, the learnable weights being learnt according to an objective function given by = r +λ d , where r is a coding cost, d is a reconstruction error, and λ is a Lagrange multiplier for controlling rate-distortion trade-off, and wherein λ is a monotonic increasing function of the popularity level such that more bits are allocated for more popular images to thereby enable the image to be compressed with the variable compression rate corresponding to the popularity level and in the negative correlation to the corresponding intrinsic popularity score.

7. The device according to claim 6 , wherein the computer program code further instructs the at least one processor to:

train the DNN based intrinsic popularity assessment model, wherein during the training the computer program code further instructs the at least one processor to:

retrieve a plurality of historical images;

construct a plurality of popularity-discriminable image pairs (PDIPs) based on the plurality of historical images, wherein each PDIP comprises a first image and a second image, the first image having a probability of higher intrinsic popularity than the second image; and

train the DNN based intrinsic popularity assessment model with the plurality of PDIPs.

8. The device according to claim 7 , wherein during the training of the DNN based intrinsic popularity assessment model with the plurality of PDIPs, the computer program code instructs the at least one processor to:

for each PDIP of the plurality of PDIPs,

determine a first intrinsic popularity score for the first image and a second intrinsic popularity score for the second image using the DNN based intrinsic popularity assessment model, and

optimise the DNN based intrinsic popularity assessment model by minimizing a binary cross entropy loss between a score difference between the first intrinsic popularity score and the second intrinsic popularity score and a ground truth binary label denoting whether the first image is intrinsically more popular than the second image.

9. The device according to claim 6 , wherein the DNN based autoencoder comprises:

a core autoencoder including an analysis transformation and a synthesis transformation to learn a quantized latent representation of a content of information; and

a hyper autoencoder arranged to learn a probabilistic model over the quantized latent representation of the content learned in the core autoencoder;

wherein the content is processed by both the core autoencoder and the hyper autoencoder to generate a compressed content of the information, and

wherein the core autoencoder and the hyper autoencoder are composed of multiple layers of conditional convolution, generalized divisive normalization (GDN), and inverse GDN, wherein the multiple layers of conditional convolution are conditioned on an intrinsic popularity score of the content of the information determined using a deep neural network (DNN) based intrinsic popularity assessment model.