Entropy Estimation Network With Group Convolution for Image Compression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image/video compression methods face challenges in accurately estimating probability distributions, which affects the size of the bitstream and overall compression performance, necessitating an efficient entropy estimation network to improve compression efficiency.
Innovation Solution
An entropy estimation method and apparatus that combines convolutional layers with pointwise operations, incorporating local attention mechanisms and group convolutions to enhance the estimation of probability distributions, utilizing a neural network structure that includes specific configurations and operations to modulate responses based on local feature matrix information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional convolutional operations are used in the entropy estimation network, then the network can process feature matrices, but the receptive field is limited and calculation complexity increases
Solution Approach 1:
The patent segments the convolutional layer into two independent branches: one performing 1x1 convolution operations and another performing 3x3 group convolution operations. This segmentation allows each branch to specialize in different aspects of feature extraction, improving the receptive field and estimation accuracy while managing calculation complexity through modular design
Solution Approach 2:
The patent transitions from traditional single-branch convolution to a multi-branch architecture operating in different spatial dimensions. The 1x1 convolution operates in the channel dimension while the 3x3 group convolution operates in the spatial dimension, creating a two-dimensional feature extraction space that expands the effective receptive field without proportionally increasing complexity
2Productivity
If the entropy estimation network uses more complex operations to improve estimation accuracy, then compression performance improves, but the bitstream size and processing overhead increase
Solution Approach 1:
The patent applies local attention mechanisms that adaptively weight different feature regions based on their local characteristics. By focusing computational resources on salient local regions rather than uniformly processing the entire feature matrix, the network achieves higher estimation accuracy for critical areas while reducing unnecessary computations in less important regions, thereby improving compression performance without proportionally increasing bitstream size
Solution Approach 2:
The patent dynamically adjusts network parameters including the group number in group convolution and attention weights based on input feature statistics. This adaptive parameter adjustment allows the network to optimize its computational complexity and estimation accuracy according to the specific characteristics of each input block, achieving better compression performance while controlling bitstream overhead through efficient parameter utilization
3Measurement precision
If group convolution is used to expand the receptive field, then more context information is captured, but the network complexity and implementation difficulty increase
Solution Approach 1:
The patent segments the feature processing into distinct 1x1 and 3x3 group convolution branches with clear functional separation. The 1x1 convolution handles channel-wise feature transformation while the 3x3 group convolution handles spatial context extraction. This segmentation provides a systematic framework that simplifies implementation by assigning specific tasks to each branch, reducing the overall implementation difficulty despite using group convolution
Solution Approach 2:
The patent implements dynamic parameter adjustment mechanisms that adapt the group convolution parameters based on input characteristics. The network can dynamically adjust the number of groups and kernel sizes according to the statistical properties of the input features, making the complex group convolution operation more adaptable and easier to implement across different scenarios while maintaining expanded receptive field benefits
Data Source
Figure 1
Figure 2
Figure 3
AI summary
This application provides an entropy estimation method and apparatus for image compression. The method includes: obtaining a to-be-processed first feature matrix, where a shape of the first feature matrix is [Cin, hin, win], Cin indicates a quantity of channels of the first feature matrix, hin indicates a height of the first feature matrix, and win indicates a width of the first feature matrix; and inputting the first feature matrix to an entropy estimation network to obtain a second feature matrix, where a shape of the second feature matrix is [Cout, hout, wout], Cout indicates a quantity of channels of the second feature matrix, hout indicates a height of the second feature matrix, and wout indicates a width of the second feature matrix. The second feature matrix and the first feature matrix meet the following conditions: Cout = Cin, hout = s · hin, and wout = s · win, where s is an integer greater than 1. The entropy estimation network includes a first convolutional layer and a pointwise operation. The first convolutional layer includes two types of convolution processing. One of the two types of convolution processing is two-dimensional group convolution. An input of the pointwise operation is respective outputs of the two types of convolution processing. In this application, compression performance can be improved.