Feature Map Video Compression for Neural Network Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing video compression methods do not effectively integrate neural network inference and image encoding/decoding, leading to inefficiencies in feature map compression and security concerns when transmitting personal information.
Innovation Solution
A method and device that simultaneously perform inference and image encoding/decoding using a neural network, involving pre-processing, feature map generation, quantization, and video conversion, with specific conditions applied to input formats and neural network layers, including convolution, graph convolution, pooling, and normalization layers, to improve coding efficiency and security.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If video compression methods are applied to feature maps, then coding efficiency is improved, but integration with neural network inference is insufficient
Solution Approach 1:
The patent merges video compression techniques with neural network inference by applying compression algorithms directly to feature maps generated during inference. This integration allows the system to perform both inference and compression in a unified framework, improving coding efficiency while maintaining adaptability to neural network operations.
Solution Approach 2:
The patent creates a multi-functional system that simultaneously performs neural network inference and video compression on feature maps. The compression method is designed to be universally applicable to different neural network architectures and inference scenarios, enhancing both productivity and adaptability.
2Reliability
If feature maps are transmitted without compression, then inference accuracy is maintained, but transmission efficiency and security are reduced
Solution Approach 1:
The patent applies compression to feature maps before transmission, performing the compression action in advance. This preliminary compression reduces the data size to be transmitted while preserving the essential information needed for accurate inference, thus improving transmission efficiency without sacrificing reliability.
Solution Approach 2:
The patent introduces compression algorithms as an intermediary process between feature map generation and transmission. This intermediary step transforms the feature maps into a compressed form that maintains inference accuracy while significantly improving transmission efficiency and enabling better security through reduced data exposure.
3Ease of operation
If personal information is transmitted in feature maps, then inference results are obtained, but security risks increase
Solution Approach 1:
The patent extracts and removes sensitive personal information from feature maps through compression techniques before transmission. By taking out the unnecessary personal details while retaining the essential feature information, the system maintains inference capability while significantly reducing security risks associated with transmitting personal data.
Solution Approach 2:
The patent converts the potential harm of transmitting personal information into a benefit by using compression to anonymize the data. The compression process inadvertently removes sensitive personal details while preserving the functional information needed for inference, thus transforming a security vulnerability into a security advantage.
Data Source
AI summary
A neural network-based image processing method and device according to embodiments of the present invention can: perform preprocessing for an input image; obtain a feature map from the pre-processed image by means of a neural network comprising a plurality of neural network layers; perform quantization for the obtained feature map; and perform video conversion with respect to the quantized feature map.