AI Encoding Device Using Integer Parameter Convolution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep Neural Networks (DNNs) require significant memory and computational resources for convolution operations, especially when handling high-resolution images, due to the use of real-number parameters, which hinders efficient encoding and decoding of images.
Innovation Solution
The proposed solution involves representing parameters of DNNs as a product of a scale factor and integer values (0 or ±2n), allowing for convolution operations to be performed using only integer multiplication and addition, reducing memory and computational demands while maintaining image quality through joint training of downscaling and upscaling DNNs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If real-number parameters are used in DNN convolution operations, then image quality is maintained, but memory and computational resources are significantly increased
Solution Approach 1:
The patent transforms DNN filter parameters from real numbers to integers by representing them as products of scale factors and integer values. This parameter transformation enables convolution operations to use only integer multiplication and addition, dramatically reducing memory requirements and computational complexity while maintaining image quality through joint training of downscaling and upscaling DNNs
Solution Approach 2:
The patent replaces the traditional real-number arithmetic system with an integer-based arithmetic system for DNN convolution operations. By substituting floating-point multiplication and addition with integer operations, the system achieves the same functional outcome with significantly reduced resource consumption
2Measurement precision
If DNN convolution operations are performed with real-number parameters, then accurate image processing is achieved, but processing time is extended
Solution Approach 1:
The patent changes the numerical representation of DNN parameters from real numbers to integers, which fundamentally alters the computational operations required. Integer-based convolution operations eliminate the need for complex floating-point arithmetic, thereby reducing processing time while maintaining processing accuracy through the joint training approach
3Manufacturing precision
If high-resolution images are processed by DNNs, then image quality is improved, but memory and calculation requirements increase significantly
Solution Approach 1:
The patent applies integer parameter representation to DNNs processing high-resolution images, transforming the computational burden. By using integer multiplication and addition instead of real-number operations, the system can handle high-resolution images with reduced memory and calculation requirements, making the processing more feasible
Data Source
AI summary
An artificial intelligence (AI) decoding apparatus includes a memory storing one or more instructions, and a processor configured to execute the stored one or more instructions to obtain image data corresponding to a first image that is downscaled from an original image by using first parameters of a first filter kernel included in a first deep neural network (DNN), reconstruct a second image corresponding to the first image, based on the obtained image data, and obtain a third image that is upscaled from the reconstructed second image, by performing an operation between the reconstructed second image and second parameters of a second filter kernel included in a second DNN corresponding to the first DNN. Each of the second parameters is represented by a product of a scale factor and one among integer values, and each of the integer values is 0 or ±2n, where n is an integer.