AI Downscaling Network for Image Transmission Quality
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image transmission systems face challenges in maintaining image quality when downscaling high-resolution images, particularly when devices receiving the images rely solely on legacy scalers, leading to deteriorated upscaling performance and limitations in using existing artificial intelligence models.
Innovation Solution
An electronic apparatus is designed with a processor and memory to learn and utilize artificial intelligence models for both downscaling and upscaling, incorporating a weighted sum of differences between various image resolutions to improve image restoration and quality, even when legacy scalers are used, allowing for optional downsampling during transmission.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If an artificial intelligence model for upscaling is not stored in the receiving device, then device complexity is reduced, but upscaling performance deteriorates
Solution Approach 1:
The system divides the image processing function between transmitter and receiver: the transmitter performs downscaling using an AI model, while the receiver uses a legacy scaler. This segmentation allows the receiver to have simpler hardware (no AI upscaling model) while still achieving good image quality through the coordinated work of both devices.
Solution Approach 2:
The patent introduces an intermediary approach where the transmitter's AI downscaling model acts as a mediator to prepare the image data, compensating for the receiver's lack of AI upscaling capability. The transmitter processes the image through AI downscaling, and the receiver simply needs to upscale using a legacy scaler, with the AI model's output serving as an intermediate that improves overall performance.
2Manufacturing precision
If existing AI models are used for image processing, then image quality is improved, but the model cannot be used when no downscaling is performed
Solution Approach 1:
The patent makes the AI model universal by enabling it to function in two modes: downscaling mode (when the transmitter has high-resolution images and wants to reduce bandwidth) and no-downscaling mode (when the transmitter wants to maintain original resolution). The model's architecture and learning parameters allow it to adapt to both scenarios, improving versatility while maintaining image quality enhancement capabilities.
3Manufacturing precision
If high-resolution images are transmitted directly, then image quality is maintained, but bandwidth consumption increases
Solution Approach 1:
The system extracts the essential information from high-resolution images by performing AI-based downscaling, which removes redundant details while preserving the most important visual characteristics. This extraction allows transmission of a compressed representation that maintains image quality perception while significantly reducing the quantity of data that needs to be transmitted over the network.
Data Source
AI summary
An electronic apparatus is disclosed. The electronic apparatus includes: a memory configured to store a downscaling network of a first artificial intelligence model, a communication interface comprising communication circuitry, and a processor connected to the memory and the communication interface and configured to control the electronic apparatus, wherein the processor is configured to: obtain an output image in which an input image is downscaled by inputting the input image the downscaling network, control the communication interface to transmit the output image to another electronic apparatus, and wherein the first artificial intelligence model is configured to be learned based on: a sample image, a first intermediate image obtained by inputting the sample image to the downscaling network, a first final image obtained by inputting the first intermediate image to an upscaling network of the first artificial intelligence model, a second intermediate image in which the sample image is downscaled by a legacy scaler, and a second final image in which the first intermediate image is upscaled by the legacy scaler.