An individual R-λ model rate control method based on super-prior variable rate image compression

Abstract

The present invention aims at the problem that the hyper-a priori variable rate image compression method cannot quickly give the trade-off parameters that the network should input according to the target bit rate, and proposes an individual R-λ model rate control method based on hyper-a priori variable rate image compression. This method first proposes an average R-λ model for the variable rate compression framework based on the hyper-a priori model by mining the relationship between the bit rate of the test sample set and the trade-off parameter λ in the distortion optimization; then, assuming that the individual R-λ model and the average R-λ model share the same structure, an individual R-λ model of adaptive individual image distribution characteristics is constructed, and on this basis, a rate control algorithm based on the individual R-λ model is established, which adjusts the encoding bit rate by adjusting the trade-off parameter λ to achieve rate control. The present invention realizes that for any image, under a given target bit rate, the trade-off parameters that should be input into the network can be quickly calculated, and the average relative error of the bit rate of the rate control is between one percent and three percent.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision, and specifically relates to technologies such as deep learning, image compression, and bit rate control. Background Art

[0002] Image compression is one of the basic topics in the field of image processing. In recent years, in the era of self-media, not only the number of pictures and videos transmitted has increased dramatically, but also the amount of data contained in a single image has continued to increase, which has put forward higher and higher requirements on the performance of image compression technology. Traditional image compression methods rely on complex manual design and individual optimization of each module, so there are more and more challenges in further pursuing improvements in coding performance. Thanks to the vigorous development of deep learning, the image compression framework based on the variational autoencoder and the model containing the hyper-prior network, that is, the hyper-prior variable bit rate image compression, has achieved better performance due to the ability to jointly optimize each module end-to-end.

[0003] Although the image can be compressed at different rates using the same model based on super-prior variable bit rate compression, the bit rates of different images are not the same when the same trade-off parameter is used as input. In practical applications, in order to adapt to the transmission bandwidth or storage requirements, it is necessary to change the input (trade-off parameter) to adjust the output bit rate of the image so that the output bit rate is close to the target bit rate, and this process requires rate control to achieve. In the traditional coding framework, the rate control algorithm usually establishes an RQ model or an R-λ model, in which the R-λ model rate control establishes a suitable RD model, uses the trade-off parameter λ in RD optimization, establishes a λ domain RD analysis, characterizes the relationship between the bit rate and the trade-off parameter, and adjusts the output bit rate by changing the trade-off parameter λ. However, due to the differences in encoders, the above-mentioned rate control method cannot be directly used in super-prior variable bit rate image compression. Summary of the invention

[0004] In view of the above problem, the present invention designs an individual R-λ model rate control method based on super a priori variable rate image compression, which can quickly calculate the trade-off parameters that should be input to the network for a target bit rate for any image, given the target bit rate.

[0005] The present invention aims at any super-a priori variable bit rate image compression and establishes an average R-λ model for a trained model.

[0006] Specifically, the designed average R-λ model is:

[0007] λ=α(e βR -1)

[0008] Among them, α and β are the parameters to be fitted. The reason for subtracting 1 in the parameter model is that when the trade-off parameter is 0, the bit rate is also 0.

[0009] Secondly, use the trained model to use 5 trade-off parameters to encode and decode each image in the Kodak dataset to obtain the bit rate, and calculate the average bit rate under each trade-off parameter, so that 5 pairs of trade-off parameters and bit rates can be obtained, and these 5 pairs of data are used to fit α and β in the average R-λ model. Note that different trained models fit different α and β. Among them, the 5 trade-off parameters must include the initial trade-off parameter λ0, and the initial trade-off parameter λ0 selects the maximum trade-off parameter λ during model training. max When one half of The remaining four trade-off parameters are optional and as dispersed as possible. In addition, when the input is the initial trade-off parameter λ0, the average bit rate is R avg0 .

[0010] Then, the individual R-λ model is designed as:

[0011]

[0012] Among them, α and β are the results of the average R-λ model fitting. ρ is the Shannon entropy R0 of the encoder output data and the average bit rate R when the initial weight parameter λ0 is avg0 The ratio of

[0013] Finally, for any input image, we first calculate the individual R-λ model, given the target bit rate R s , the individual R-λ model of the image can be used to calculate the trade-off parameter λ that should be fed into the network s , the bit rate output by the encoder will be very close to the target bit rate, thus achieving bit rate control.

[0014] Beneficial Effects

[0015] The present invention is the first to implement a rate control method based on super-a priori variable rate image compression. In order to verify the effectiveness of the present invention, three different super-a priori variable rate image compression models were tested, and the average relative error of the rate control of the present invention was between 1% and 3%. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 , a flow chart of the bit rate control method of the present invention; DETAILED DESCRIPTION

[0017] First, given a trained variable bit rate compression model, five trade-off parameters are input into the network to encode and decode each image in the Kodak dataset to obtain the bit rate, and the average bit rate under each trade-off parameter is calculated. The five trade-off parameters must include the initial trade-off parameter λ0, and the initial trade-off parameter λ0 selects the maximum trade-off parameter λ during model training. max When one half of The remaining parameters are λ max In addition, these five pairs of trade-off parameters and code rate data are substituted into λ=α(e βR -1), and fit α and β. And additionally record the average bit rate R when the initial trade-off parameter λ0 avg0 The value of .

[0018] Then, for any input image, the initial weight parameter λ0 is fed into the network to obtain the Shannon entropy R0 of the encoder output data, which is calculated by Get ρ. At this point, α, β and ρ in the individual R-λ model are all known.

[0019] Finally, for a given target bit rate R target , brought into the individual R-λ model, we can obtain the trade-off parameters that should be fed into the network At this point, the bit rate output by the network will be very close to the target bit rate, thus achieving bit rate control.

[0020] The present invention is the first to implement a rate control method based on a super-a priori variable rate compression network. In order to verify the effectiveness of the present invention, three different super-a priori variable rate compression networks were tested, and the average relative error of the rate control was between 1% and 3%.

Claims

1. An individual R-λ model rate control method based on super-a priori variable rate image compression, characterized in that: For any image, the trained image compression model is used to encode and decode multiple images using at least 5 trade-off parameters to obtain the bit rate, and the average bit rate under each trade-off parameter is calculated; each trade-off parameter and the corresponding average bit rate constitute a set of data, and multiple sets of data are used to fit the parameters α and β in the average R-λ model; the fitted parameters α and β are substituted into the individual R-λ model, and the target bit rate R can be obtained according to the given target bit rate R. s Calculate the trade-off parameter λ corresponding to the input image that should be fed into the image compression model s , to achieve bit rate control; the average R-λ model is: Among them, α and β are the parameters to be fitted, λ is the trade-off parameter, and R is the bit rate; The individual R-λ model is designed as: Send the initial weight parameter λ0 into the network to obtain the Shannon entropy of the encoder output data Where R avg0 is the average bit rate.

2. The individual R-λ model rate control method based on super a priori variable rate image compression according to claim 1, characterized in that: Different α and β are fitted using different trained image compression models.

3. The individual R-λ model rate control method based on super a priori variable rate image compression according to claim 1, characterized in that: The trade-off parameters include the initial trade-off parameter λ0, and when the image compression model is trained, The remaining four trade-off parameters are optional.

Images