Lossless HDR Image Inferencing via Bit Partitioning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Edge devices face challenges in performing accurate inferencing operations on high dynamic range (HDR) images due to the need to quantize 24-bit pixel values to 8-bit operands, resulting in loss of information and precision, which leads to inaccurate results.
Innovation Solution
The method involves subdividing the K bits of pixel data into M partitions, where each partition corresponds to the N-bit operand size of the AI accelerator, allowing parallel processing across multiple channels, thereby enabling lossless inferencing operations without quantization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If 24-bit pixel values are quantized to 8-bit operands to match AI accelerator capabilities, then the device complexity and processing efficiency are improved, but the measurement precision and information completeness deteriorate
Solution Approach 1:
The patent divides each 24-bit pixel value into multiple 8-bit segments (typically three segments: R, G, and B channels with 8 bits each). This segmentation allows the AI accelerator to process each segment separately using its native 8-bit operand width, avoiding quantization loss while maintaining processing efficiency. The segmented values are then recombined to reconstruct the full 24-bit pixel value for output.
2Ease of operation
If 24-bit pixel values are quantized to 8-bit operands, then the ease of operation with AI accelerators is improved, but the reliability of inferencing results deteriorates
Solution Approach 1:
By segmenting the 24-bit pixel value into three separate 8-bit channels, the patent enables direct compatibility with AI accelerators that operate on 8-bit operands without requiring quantization. Each segment is processed independently through the neural network, preserving the full information content and ensuring reliable inferencing results.
Data Source
AI summary
This disclosure provides methods, devices, and systems for neural network inferencing. The present implementations more specifically relate to performing inferencing operations on high dynamic range (HDR) image data in a lossless manner. In some aspects, a machine learning system may receive a number (K) of bits of pixel data associated with an input image and subdivide the K bits into a number (M) of partitions based on a number (N) of bits in each operand operated on by an artificial intelligence (AI) accelerator, where N<K. For example, the K bits may represent a pixel value associated with the input image. In some implementations, the AI accelerator may perform an inferencing operation based on a neural network by processing the M partitions, in parallel, as data associated with M channels, respectively, of the input image.