Block-encoding raster images with separable kernels
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image encoding methods for CMOS sensors face challenges such as limited transmission rates, high power consumption, and memory constraints, particularly in line-by-line sequential read modes, which hinder efficient compression and the use of complex filtering algorithms in consumer devices.
Innovation Solution
A method for block-encoding raster images using successive two-dimensional decompositions with one-dimensional kernels for vertical and horizontal decompositions, optimizing pixel block sizes and decomposition levels to enhance consistency and reduce memory usage, combined with entropic transcoding for efficient data representation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of energy
If traditional block-encoding methods (e.g., DCT, wavelets) are applied to square blocks of pixels, then image compression is achieved, but memory usage increases and power consumption rises due to the need to store and process entire blocks before decomposition
Solution Approach 1:
The patent divides the image into strips of N lines that are processed sequentially, rather than loading entire blocks into memory. Each strip is decomposed line-by-line using separable kernels, allowing compression without storing complete blocks in memory simultaneously. This segmentation reduces memory requirements while maintaining compression effectiveness.
Solution Approach 2:
The patent applies preliminary separable decomposition using one-dimensional kernels in the horizontal and vertical directions before final encoding. This preliminary action transforms the data into a format that requires less memory for subsequent processing, enabling efficient compression with reduced power consumption in line-by-line read modes.
2Reliability
If complex filtering algorithms are implemented in image capture devices, then image quality improves, but transmission rate limits and memory constraints prevent their efficient use
Solution Approach 1:
The patent changes the processing parameters by using separable kernels and line-by-line decomposition, transforming the computational requirements into a format that fits within transmission rate limits. This allows complex filtering to be applied efficiently during compression without exceeding bandwidth constraints.
Solution Approach 2:
The patent replaces traditional block-based mechanical processing with a line-by-line sequential processing system using separable mathematical kernels. This substitution enables complex filtering algorithms to run efficiently within the constraints of limited transmission rates and memory capacity.
3Use of energy by stationary object
If line-by-line sequential read mode is used for image acquisition, then power consumption is reduced, but existing encoding methods cannot efficiently compress the data due to lack of contextual information from adjacent lines
Solution Approach 1:
The patent segments the image processing into N-line strips that are processed sequentially in line-by-line mode. By using separable kernels, the method maintains compression efficiency while adapting to the sequential read mode, allowing contextual information to be utilized across lines without requiring simultaneous memory storage of entire blocks.
Solution Approach 2:
The patent creates a universal encoding method that works with line-by-line sequential read modes while maintaining compression efficiency. The separable kernel approach provides multi-functionality, enabling the system to achieve both low power consumption and high compression efficiency by processing data in a format that matches the sequential acquisition mode.
Data Source
AI summary
A method of block-encoding a raster image by successive two-dimensional decompositions of blocks of the image in a base of functions using a combined application of a one-dimensional kernel of vertical decomposition of n pixels and of a one-dimensional kernel of horizontal decomposition of p pixels. In the method the horizontal dimension P of each block is determined as a multiple of p, P=k·p, and a decomposition at logp(P) level(s) of resolution is accomplished using the horizontal decomposition kernel, and the vertical dimension N of each block is determined as a multiple of n, N=l·n, and a decomposition at logp(N) level(s) of resolution is accomplished using the vertical decomposition kernel. For given values of n and p, the values of k and l are chosen such that the vertical dimension N is strictly less than the horizontal dimension P.