AI Image Encoding Using Downscale Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image encoding and decoding methods are inefficient for high-resolution images, leading to increased bitrate and computational demands, particularly in video streaming services, where the amount of information to be processed grows rapidly with image resolution.
Innovation Solution
The implementation of an artificial intelligence (AI) encoding and decoding system using deep neural networks (DNNs) for downscaling and upsampling images, where the AI downscale process generates a reduced image and a residual image, which are then processed through a DNN to reduce bitrate and improve encoding efficiency, and the AI upscale process reconstructs the original image quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional encoding methods are used for high-resolution images, then image quality is maintained, but bitrate and computational demands increase significantly
Solution Approach 1:
The patent segments the encoding process into two distinct stages: a training stage where a neural network learns optimal compression representations, and an encoding stage where the trained network compresses images. This segmentation allows the system to achieve high compression ratios during actual encoding while maintaining quality, as the neural network has pre-learned efficient representations during training.
Solution Approach 2:
The patent performs preliminary training of the neural network encoder-decoder pair before actual image encoding. During this preliminary action, the network learns optimal compression strategies and representations. When actual encoding occurs, the pre-trained network can efficiently compress images with high quality retention, avoiding the need for complex real-time optimization during encoding.
2Measurement precision
If conventional encoding methods are used for high-resolution images, then image quality is maintained, but computational complexity increases
Solution Approach 1:
The patent divides the computational workload into training phase computations and encoding phase computations. The complex neural network architecture and weight optimization occur during training, while the actual encoding uses the pre-trained network which performs simpler forward propagation operations. This segmentation reduces real-time encoding complexity.
Solution Approach 2:
The patent replaces traditional mechanical encoding algorithms (such as discrete cosine transform and motion compensation in conventional video codecs) with a neural network-based system. The neural network learns optimal compression transformations data-driven, substituting fixed mathematical transforms with adaptive learned representations that can be more efficient for specific image types.
3Measurement precision
If image resolution is increased for high-quality display, then image quality improves, but processing time and computational resources increase
Solution Approach 1:
The neural network is pre-trained on large datasets of high-resolution images to learn efficient compression and representation strategies. During actual processing, the pre-trained network can quickly encode and decode images without requiring complex real-time computations, thus reducing processing time while maintaining quality.
Solution Approach 2:
The patent changes the parameters of the encoding system by using learned neural network weights instead of fixed encoding parameters. The network can adapt its internal parameters based on the input image characteristics, allowing efficient processing of high-resolution images by learning task-specific optimizations during training rather than using generic high-complexity encoding for all images.
Data Source
AI summary
An artificial intelligence (AI) encoding apparatus includes a processor configured to execute one or more instructions stored in the AI encoding apparatus to: input, to a downscale deep neural network (DNN), a first reduced image downscaled from an original image and a reduction feature map having a resolution lower than a resolution of the original image; obtain a first image AI-downscaled from the original image in the downscale DNN; generate image data by performing a first encoding process on the first image; and output the image data.