Neural Feature Matrix Edge Compression for Bandwidth-Efficient Encoding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning feature encoding for machine vision is hindered by bandwidth waste and reduced transmission efficiency due to padding operations in convolutional neural networks, which generate numerous feature values at edges that are transmitted to decoders.
Innovation Solution
An encoder identifies and compresses rows and columns with identical feature values at the edges of feature matrices from intermediate layers, deleting these values and encoding the remaining values and edge values, along with the number of compressed rows and columns, to improve transmission efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If padding operations are performed in convolutional neural networks to add pixels to the periphery of images, then the neural network can process edge features, but bandwidth waste and reduced transmission efficiency occur due to numerous edge feature values being transmitted to decoders
Solution Approach 1:
The patent extracts and identifies edge rows and columns from the feature matrix that contain redundant padding information. By separating these edge elements from the main feature data, the system can selectively transmit only the necessary feature values to the decoder, eliminating bandwidth waste while preserving edge feature processing capability.
Solution Approach 2:
The patent applies different processing strategies to different regions of the feature matrix. Edge regions are identified and handled separately from internal regions, with edge rows and columns being compressed or omitted in transmission. This local differentiation allows the system to maintain edge feature capability where needed while reducing transmission overhead where padding redundancy exists.
2Loss of information
If all feature values including edge values are transmitted to the decoder, then complete feature information is preserved, but transmission latency and bandwidth consumption increase
Solution Approach 1:
The encoder performs preliminary identification and marking of edge rows and columns before transmission. By pre-processing the feature matrix to identify which elements are edge padding versus meaningful features, the system can optimize the transmission data structure in advance, reducing both the amount of data to be transmitted and the time required for transmission without losing critical feature information.
Solution Approach 2:
The patent transmits a partial set of feature values by selectively omitting or compressing edge region data. Instead of transmitting all feature values, the system transmits only the essential internal feature values and selectively included edge values, achieving sufficient feature completeness for decoder reconstruction while significantly reducing transmission latency and bandwidth consumption.
3Loss of energy
If edge feature values are compressed and transmitted, then bandwidth efficiency is improved, but the complexity of identifying and processing edge rows and columns increases
Solution Approach 1:
The patent segments the feature matrix into distinct regions: edge rows, edge columns, and internal regions. By dividing the feature space into these segments, the system can apply different processing rules to each segment. The segmentation is performed through systematic identification of boundary rows and columns, which, while adding some complexity, provides a structured approach that simplifies the overall compression strategy and enables efficient bandwidth utilization.
Data Source
AI summary
The method of the present disclosure includes: for a feature matrix of each channel of an image output from an intermediate layer of a neural network: determining row(s) and column(s) having same feature values at edges of the feature matrix as row(s) to be compressed and column(s) to be compressed, deleting the feature values of the row(s) to be compressed and the column(s) to be compressed, and reserving remaining feature values as reserved values; compressing the feature values of the row(s) to be compressed and the column(s) to be compressed to obtain edge value(s); and encoding the reserved values and the edge value(s), as well as a number of the row(s) to be compressed and a number of the column(s) to be compressed, and sending a result of the encoding to a decoder for restoring the feature matrix of the each channel by the decoder.