Hierarchical Encoder for AI Image Quantization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing systems face challenges in efficiently encoding and transmitting image data for AI recognition, particularly when imaging devices have specification constraints that limit the ability to set different quantization values for various areas within an image, leading to reduced data size reduction and usability of non-target areas in decoded data.
Innovation Solution
An image processing system that employs a hierarchical encoder to determine and encode target and non-target areas with specific quantization values, generating separate encoded data streams, and a transcoder that reconstructs and re-encodes the data to maintain high recognition accuracy and reduce data volume, allowing the use of non-target areas as image data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If the same quantization value is set for all areas in an image, then the imaging device can operate within specification constraints, but the data size reduction is reduced and non-target areas cannot be effectively used
Solution Approach 1:
The patent segments the image into target areas and non-target areas, and applies different quantization values to each segment. The hierarchical encoder determines target areas based on AI recognition needs and applies higher quantization values to preserve detail where needed, while using lower quantization values for non-target areas to reduce data size. This segmentation approach resolves the contradiction by allowing differentiated processing within the same image while maintaining compliance with overall system specifications.
Solution Approach 2:
The patent implements local quality by applying different quantization values to different spatial locations within the image. Target areas receive higher quantization values to maintain recognition accuracy, while non-target areas receive lower quantization values to minimize data consumption. This local differentiation enables the system to optimize both recognition performance and data efficiency simultaneously, resolving the contradiction between specification compliance and data size reduction.
2Quantity of substance
If different quantization values are set for each area, then data size reduction is improved, but the imaging device specification constraints are violated
Solution Approach 1:
The patent introduces dynamic quantization value adjustment based on the hierarchical encoder's analysis of target and non-target areas. The system dynamically determines appropriate quantization values for different regions rather than using a fixed approach. This dynamic adaptation allows the system to achieve optimal data size reduction while maintaining compliance with imaging device specifications through intelligent, context-aware quantization selection.
Solution Approach 2:
The hierarchical encoder acts as an intermediary between the imaging device and the encoding process. It analyzes the image, identifies target areas, and determines appropriate quantization values before encoding occurs. This intermediary role enables the system to bridge the contradiction by translating specification constraints into optimized encoding parameters that achieve both data size reduction and compliance.
3Quantity of substance
If black-painting is performed on non-target areas, then data size is reduced, but the non-target areas become unusable for image data purposes
Solution Approach 1:
The patent applies parameter changes by adjusting quantization values rather than using black-painting to handle non-target areas. By applying lower quantization values to non-target areas, the system reduces data size while preserving the visual information and usability of these areas. This parameter-based approach eliminates the need for black-painting, thereby maintaining data size reduction benefits while preserving the versatility and usability of non-target areas for subsequent image processing tasks.
Data Source
AI summary
An image processing system includes a hierarchical encoder that determines, based on recognition processing, a target area needed to recognize a recognition target and a non-target area other than the target area in image data, a quantization value of the target area needed to recognize the recognition target, and a quantization value of the non-target area, encodes an entire area of the image data with the quantization value of the target area to generate first encoded data, and encodes the entire area of the image data with the quantization value of the non-target area to generate second encoded data, and a transcoder that generates reconstructed image data by using the target area in first decoded data obtained by decoding the first encoded data and the non-target area in second decoded data obtained by decoding the second encoded data, and re-encodes the reconstructed image data to generate re-encoded data.