High-Parallel Autoregressive Scanning With Masked Convolution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing autoregressive scanning methods in digital image processing struggle to achieve a balance between high parallelism and performance, particularly in large-resolution images, leading to inefficiencies in image generation and compression tasks.
Innovation Solution
A probability modeling-oriented high-parallel autoregressive scanning and masked convolution design method that establishes a mathematical relationship between scanning and image resolution, constructs specific scanning angles, and employs masked convolutions to enhance parallelism and performance through advanced indexing and training with cross-entropy loss.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If wavefront scanning is used to increase parallelism, then the number of scanning steps is reduced, but the convolution kernel size must be limited which reduces modeling accuracy
Solution Approach 1:
The patent transforms the traditional 1D sequential scanning into 2D block-based parallel scanning. By dividing the image into multiple blocks and scanning them in parallel with different offsets, the method achieves both high parallelism and large convolution kernel support. Each block can be processed independently with its own convolution kernel, allowing large kernels without increasing total scanning steps.
2Productivity
If diagonal scanning is used to increase parallelism, then scanning efficiency improves, but the valid receptive field is concentrated in one area limiting performance
Solution Approach 1:
The patent segments the image into multiple blocks that can be scanned in parallel. Each block maintains its own valid receptive field through offset-based indexing, ensuring comprehensive coverage across the entire image. This segmentation allows different parts of the image to be processed simultaneously while each retains full contextual information.
3Manufacturing precision
If serial scanning is used to ensure modeling accuracy, then performance is maintained, but the scanning process cannot be parallelized reducing efficiency
Solution Approach 1:
The patent performs preliminary block division and offset assignment before the actual scanning process. By pre-calculating the scanning offsets and block assignments, the system enables parallel execution of multiple blocks while maintaining the same modeling accuracy as sequential scanning. The preliminary organization of data allows subsequent parallel processing without compromising accuracy.
Data Source
AI summary
A probability modeling-oriented high-parallel autoregressive scanning and masked convolution design method is provided. The method includes establishing a mathematical relationship between a number of scanning and an image resolution when the number of scanning and the image resolution exhibit a linear relationship; and constructing a specific scanning angle to meet a limit of a given number of scanning. The method also includes constructing a mask mode of masked convolution under the specific scanning angle. Compared with wavefront scanning, the number of scanning of the scanning mode according to the method is not related to the size of the convolution kernel, but related to the mask mode, so that the larger convolution kernel can be used to enhance the probability estimation performance of the model under the condition that the number of scanning is unchanged.