Feature Map Compression via Selective Entropy Encoding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for compressing feature maps in neural networks do not efficiently address spatial-temporal redundancy and inter-view redundancy, leading to suboptimal coding efficiency.
Innovation Solution
A method and device for compressing feature tensors using skip coding and partition structures with various hierarchical structures, which selectively applies entropy encoding based on the neural network structure and layer types.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If entropy encoding is applied to all symbols in the feature tensor, then coding completeness is improved, but coding efficiency deteriorates due to processing overhead
Solution Approach 1:
The feature tensor is segmented into multiple channels, and entropy encoding is selectively applied to specific channels based on their characteristics. This allows the system to process only the necessary portions of the data through entropy encoding, reducing overall processing overhead while maintaining coding completeness for channels that require it.
Solution Approach 2:
Different encoding strategies are applied to different channels based on their local characteristics. Channels with higher information content or greater variability undergo entropy encoding, while channels with lower information content use simpler encoding methods, optimizing the balance between coding completeness and efficiency.
2Productivity
If skip coding is applied to reduce processing overhead, then coding efficiency is improved, but coding precision deteriorates due to selective entropy encoding
Solution Approach 1:
The system applies different encoding precisions to different channels based on their local characteristics. Channels that require higher precision for accurate representation undergo full entropy encoding, while channels with lower precision requirements use skip coding, thus maintaining overall coding precision where needed while improving efficiency elsewhere.
Solution Approach 2:
The encoding strategy is dynamically adjusted based on channel characteristics and content importance. The system can adaptively determine which channels require entropy encoding versus skip coding, allowing flexible optimization between precision and efficiency based on the specific features of the input data.
3Device complexity
If feature tensor is processed without partitioning, then processing simplicity is improved, but redundancy removal efficiency deteriorates
Solution Approach 1:
The feature tensor is partitioned into multiple channels, enabling independent processing of each channel. This segmentation allows the system to identify and remove redundancies within and between channels more effectively, as each channel can be analyzed and compressed separately based on its specific characteristics.
Solution Approach 2:
The partitioning introduces a channel dimension for independent processing, allowing redundancy removal to operate effectively across different dimensions of the feature tensor. This multi-dimensional approach enables more thorough redundancy elimination compared to processing the entire tensor as a single unit.
Data Source
AI summary
A neural network-based image processing method and apparatus according to an embodiment of the present invention may: acquire a feature tensor from an input image by using a first neural network including a plurality of neural network layers; acquire a symbol tensor by performing quantization on the acquired feature tensor; and generate a bitstream by performing entropy encoding on the basis of the symbol tensor.