CN-120639976-B - Depth learning image compression method and device capable of automatically controlling code rate and uniformly coding

CN120639976BCN 120639976 BCN120639976 BCN 120639976BCN-120639976-B

Abstract

The invention provides a depth learning image compression method and device capable of automatically controlling code rate and uniformly coding, wherein the method comprises the following steps: inputting the images or the characteristics into a built trained compression model containing a variable component self-encoder, and outputting the compressed images or the characteristics after the compression optimization of the input images or the characteristics by implementing the characteristic channel uniform grouping and the dynamic control code rate. The method ensures that the sub-packet bytes are uniform through uniform coding, controls the total sub-packet bytes in a specified range through automatic code rate control, and can realize accurate control of the occupation of the compressed bytes by combining the sub-packet bytes with the specific limit of bandwidth and transmission rate in the process of satellite or internet transmission scenes with limited bandwidth, thereby ensuring the maximum transmission efficiency.

Inventors

Yi Zhongli
LI HUIMIN
ZHENG XINRUI
CHEN PENGFEI
ZHU GUIDONG
WANG FUZHAI
WANG PANPAN
WANG WEICHANG

Assignees

交通运输部规划研究院
广州海格星航信息科技有限公司

Dates

Publication Date: 20260508
Application Date: 20250626

Claims (6)

1. A method for automatically controlling code rate and uniformly encoding depth learning image compression, the method comprising: inputting images or features into a constructed trained compression model containing a variable self-encoder; The compression model performs compression optimization on an input image or feature by implementing even grouping of feature channels and dynamic control of code rate, and outputs the compressed image or feature, and the method comprises the following steps: The compression model uniformly groups the model channels into equal-width sub-channels according to the total number of the model channels and the number of the sub-packages, wherein each equal-width sub-channel comprises the same number of characteristic channels; Each equal-width sub-channel code generates an independent compressed sub-packet; Calculating the standard deviation of the number of bits among the sub-packets, adding the standard deviation of the number of bits among the sub-packets as a penalty term into a loss function, and automatically correcting the volume deviation of the sub-packets through gradient descent; training of compression model dynamic rate control, comprising: presetting a target code rate range; model preheating is carried out by adopting fixed super parameters in a preset number of training periods; Starting the next training period of the preset number of training periods, monitoring the current code rate in real time, automatically adjusting the super-parameters when the current code rate exceeds the target code rate range, and injecting the updated super-parameters into the loss function in real time; stabilizing the code rate within the target code rate range through iterative adjustment; the expression of the loss function is as follows: In the above formula, D is the mean square error loss of the original image and the reconstructed image, and measures the quality of the compressed image, R is the compression code rate, and the compression degree is generally measured by the bit number occupied by a single pixel after bpp compression; , ,..., Representing the independent compression code rate of 1-n channels, and controlling the balance of the compression code rate and the image quality by the super-parameter lambda.
2. The method according to claim 1, wherein the dynamic code rate control is adjusted by: when the current code rate is detected to be lower than the lower limit of the target code rate range, increasing the super-parameters; When the current code rate is detected to be higher than the upper limit of the target code rate range, reducing the super-parameters; and the parameter adjustment amplitude is dynamically calculated according to the code rate deviation proportion.
3. The method of claim 1, wherein the training of the compression model dynamic rate control further comprises: setting a super-parameter value range, and cutting off according to the nearest boundary value when the adjusted super-parameter value exceeds the super-parameter value range.
4. A depth learning image compression apparatus for automatically controlling a code rate and uniformly encoding, the apparatus comprising: The image characteristic input module is used for inputting images or characteristics into the constructed trained compression model containing the variational self-encoder; the compression output module is used for carrying out uniform grouping and dynamic control of code rate on the input images or features by the compression model, and outputting the compressed images or features after compression optimization; the compression output module includes: the feature channel uniform grouping sub-module is used for uniformly grouping the model channels into equal-width sub-channels according to the total number of the model channels and the number of the sub-packages by the compression model; the coding sub-module is used for generating independent compressed sub-packets by coding each equal-width sub-channel; the device also comprises a uniform coding training module, wherein the uniform coding training module is used for: Calculating the standard deviation of the number of bits among the sub-packets, adding the standard deviation of the number of bits among the sub-packets as a penalty term into a loss function, and automatically correcting the volume deviation of the sub-packets through gradient descent; The device also comprises a dynamic code rate control training module, wherein the dynamic code rate control training module is used for: presetting a target code rate range; model preheating is carried out by adopting fixed super parameters in a preset number of training periods; Starting the next training period of the preset number of training periods, monitoring the current code rate in real time, automatically adjusting the super-parameters when the current code rate exceeds the target code rate range, and injecting the updated super-parameters into the loss function in real time; stabilizing the code rate within the target code rate range through iterative adjustment; the expression of the loss function is as follows: In the above formula, D is the mean square error loss of the original image and the reconstructed image, and measures the quality of the compressed image, R is the compression code rate, and the compression degree is generally measured by the bit number occupied by a single pixel after bpp compression; , ,..., Representing the independent compression code rate of 1-n channels, and controlling the balance of the compression code rate and the image quality by the super-parameter lambda.
5. An electronic device comprising a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus; The memory is used for storing a computer program; The processor is configured to implement the method for compressing a deep learning image with automatic code rate control and uniform coding according to any one of claims 1 to 3 when executing a program stored in a memory.
6. One or more computer-readable media having instructions stored thereon that, when executed by one or more processors, cause the processors to perform the automatically controlled code rate and uniformly encoded deep learning image compression method of any of claims 1-3.

Description

Depth learning image compression method and device capable of automatically controlling code rate and uniformly coding Technical Field The present invention relates to the field of image processing technology, and in particular, to a method for compressing a depth learning image with automatic code rate control and uniform coding, a device for compressing a depth learning image with automatic code rate control and uniform coding, an electronic device, and a computer readable medium. Background Conventional image progressive compression algorithms are mainly based on wavelet transforms and bit-plane coding. The progressive image compression algorithm based on wavelet transformation utilizes wavelet transformation to decompose images into wavelet coefficients with different scales and directions, and the method can generate artifacts such as ringing effect and the like when processing complex image structures and textures, thereby influencing the image quality. The bit plane coding method is to code each bit plane of the binary representation of the image pixel, and code and transmit from the most significant bit, but the method has insufficient utilization of correlation between different bit planes, and the compression efficiency needs to be improved. Image compression algorithms based on deep learning are rapidly developed and can be broadly divided into methods based on variational self-encoders (VAEs), based on generating countermeasure networks (GAN), and based on self-supervised learning. The variational self-encoder (VAE) maps the image to the probability distribution of the potential space for encoding and decoding, has certain advantages in progressive compression, can process the main part of the potential space to obtain a rough image, and then gradually process other parts to improve the image quality, but has higher model training and calculation complexity. The generated countermeasure network (GAN) consists of a generator and a discriminator, wherein in progressive image compression, the generator reconstructs the compressed representation into a high-quality image, the discriminator judges the similarity between the reconstructed image and the original image, and the reconstruction capability of the generator is optimized through countermeasure training, but the training process is unstable, and the problems of collapse and the like are easy to occur. The self-supervision learning method enables the model to learn useful characteristic representations from the data, and in progressive image compression, the characteristics obtained by self-supervision learning are utilized to construct preliminary image representations first, and then the image quality is further improved, but the adaptability in large-scale image data application is required to be improved. The conventional image compression algorithm based on deep learning and the conventional image compression algorithm based on deep learning are both fixed code rates, the code rates cannot be controlled in a specified range accurately, and the size difference among sub-packets which are decompressed gradually is large, so that the transmission efficiency is difficult to maximize in a satellite or internet transmission scene with limited transmission bandwidth. Disclosure of Invention The present invention has been made in view of the above problems, and has as its object to provide an automatic code rate controlling and uniformly encoded deep learning image compression method and corresponding automatic code rate controlling and uniformly encoded deep learning image compression device, an electronic apparatus, and a computer readable medium, which overcome or at least partially solve the above problems. The invention discloses a depth learning image compression method capable of automatically controlling code rate and uniformly coding, which comprises the following steps: inputting images or features into a constructed trained compression model containing a variable self-encoder; the compression model performs compression optimization on the input images or features by implementing even grouping of feature channels and dynamically controlling code rate, and then outputs the compressed images or features. Optionally, the compression model performs compression optimization on an input image or feature by implementing feature channel uniform grouping and dynamic control of code rate, and then outputs the compressed image or feature, which includes: The compression model uniformly groups the model channels into equal-width sub-channels according to the total number of the model channels and the number of the sub-packages, wherein each equal-width sub-channel comprises the same number of characteristic channels; Each equal width sub-channel code generates an independent compressed sub-packet. Optionally, the training of the feature channel uniform grouping strategy to equalize the sub-packet volumes includes: calculating the standard deviation