Neural Network Dequantization for Signal Estimation Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for extending the number of quantization bits in signals fail to accurately reflect the fine information of the original signal, as they rely solely on pre-extended digital signal information, neglecting the original characteristics of the signal.
Innovation Solution
A neural network is trained using learning data that includes low-bit and high-bit quantized signals to estimate a high-bit output signal by adding a signal from the output layer to the low-bit input signal, utilizing a multilayer structure with an input layer and output layer, and employing techniques like gated convolutional neural networks or attention structures.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional methods (FIR/IIR filters, spline interpolation, simulated annealing) are used to extend quantization bits, then the extension process can be completed using only pre-extended digital signal information, but the fine information of the original signal is not reflected in the estimation result
Solution Approach 1:
The patent applies preliminary action by pre-processing the low-bit signal to extract residual information before feeding it into the neural network. The residual signal, which contains fine information lost during quantization, is generated in advance and used as input to guide the neural network's estimation process, thereby preserving fine information that would otherwise be lost in traditional extension methods
Solution Approach 2:
The patent introduces a residual signal as an intermediary between the low-bit input signal and the high-bit estimation process. This residual signal acts as a mediator that carries fine information from the original signal, allowing the neural network to incorporate this information into its estimation without directly accessing the original high-bit signal
2Measurement precision
If a neural network with multilayer structure is used to process low-bit signals, then fine information can be reflected in the estimation result, but the device complexity increases
Solution Approach 1:
The patent segments the signal processing task into distinct components: residual signal generation, neural network processing, and signal reconstruction. The neural network itself is segmented into multiple layers (input layer, hidden layers, output layer), each performing specific functions. This segmentation allows the complex task of high-bit signal estimation to be broken down into manageable parts that can be optimized independently
Solution Approach 2:
The patent replaces traditional mechanical signal processing methods (FIR/IIR filters, spline interpolation, simulated annealing) with a neural network-based system. This substitution transitions from deterministic mathematical operations to a learning-based approach that can automatically adapt to the characteristics of the input signal, achieving better performance despite increased computational complexity
Data Source
AI summary
A neural network is learned that uses learning data including a low-bit signal obtained by quantizing a signal to a first number of quantization bits and a high-bit signal obtained by quantizing the signal to a second number of quantization bits larger than the first number of quantization bits, to receive as an input a low-bit input signal obtained by quantizing an input signal to the first number of quantization bits and output an estimated signal of a high-bit output signal obtained by quantizing the input signal to the second number of quantization bits. This neural network has a multilayer structure including an input layer and an output layer, and obtains and outputs an estimated signal of a high-bit output signal obtained by adding to a low-bit input signal a signal output from the output layer in response to the low-bit input signal being input to the input layer.