Neural Network Feature Quantization for Image Bitstreams
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image encoding/decoding technologies face challenges in achieving efficient encoding/decoding, feature quantization/dequantization, and signaling of feature quantization-related information, especially for high-resolution and high-quality images.
Innovation Solution
The proposed solution involves an image encoding/decoding method and apparatus that utilize an artificial neural network-based feature extraction method. This approach includes obtaining a feature set from an input image, determining the importance of each channel, and performing feature quantization/dequantization based on the importance of the channels, with the number of quantization bits determined accordingly.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If high-resolution and high-quality image data are transmitted, then image quality is improved, but transmission cost and storage cost increase
Solution Approach 1:
The patent applies local quality by differentiating the treatment of different feature channels. Important channels (such as those containing edge, texture, or semantic information) are quantized with higher precision using more bits, while less important channels are quantized with lower precision using fewer bits. This selective quantization strategy preserves critical image features while reducing the overall bitstream size, thereby maintaining image quality while降低 transmission and storage costs.
2Measurement precision
If feature quantization is performed with high precision, then decoding accuracy is improved, but transmission cost increases
Solution Approach 1:
The patent dynamically changes the quantization parameter (number of bits) based on the importance of each feature channel. The encoder determines channel importance metrics and adjusts the quantization precision accordingly, allocating more bits to important channels and fewer bits to less important channels. This adaptive parameter adjustment optimizes the balance between decoding accuracy and transmission efficiency.
3Productivity
If channel-wise quantization is performed, then encoding efficiency is improved, but signaling complexity increases
Solution Approach 1:
The patent segments the feature channels into different groups based on their importance characteristics. Instead of individually signaling quantization parameters for each channel, the encoder groups channels with similar importance levels and applies uniform quantization parameters to each group. This segmentation strategy reduces the number of signaling operations required while maintaining the benefits of channel-wise adaptive quantization.
Data Source
AI summary
An image encoding/decoding method and apparatus are provided. An image decoding method may comprise obtaining, from a bitstream, a feature set extracted from an input image using an artificial neural network-based feature extraction method, first information on importance of each of a plurality of channels included in the feature set and second information on a quantization method of the feature set, decoding the feature set, the first information and the second information, and dequantizing the decoded feature set based on the decoded first information and the decoded second information. The second information may comprise the number of quantization bits of each of the plurality of channels, and the number of quantization bits may be determined based on the importance of each of the plurality of channels.