Hardware-Implemented Quantized CNN for Real-Time Video Scoring
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Convolutional Neural Networks (CNNs) face challenges in computational intensity and resource inefficiency due to extensive parameters and floating-point operations, limiting their practicality in real-time video processing tasks.
Innovation Solution
A hardware-implemented quantized CNN design that avoids floating-point operations, using quantization techniques and binary shift operations, optimized for efficiency and reduced resource usage, with a custom silicon-based integrated circuit for real-time video quality assessment.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If CNNs use extensive parameters and floating-point operations to maintain accuracy, then measurement precision is improved, but computational intensity and resource consumption increase
Solution Approach 1:
The patent applies quantization to change the parameter representation from floating-point to fixed-point or integer formats. This reduces the precision requirement for internal computations while maintaining sufficient accuracy for video quality assessment, thereby decreasing computational resource consumption and energy usage.
Solution Approach 2:
The patent uses a simplified hardware architecture that copies the essential computational functionality of traditional CNNs but implements it with resource-efficient operations. The hardware design replicates the neural network's processing capability while using minimal logic gates and memory resources, achieving accurate video quality assessment with reduced computational intensity.
2Measurement precision
If CNNs use floating-point operations to ensure computational accuracy, then measurement precision is improved, but device complexity and hardware footprint increase
Solution Approach 1:
The patent replaces the complex floating-point arithmetic hardware with simpler fixed-point or integer arithmetic units. This substitution eliminates the need for complex floating-point dividers, multipliers, and normalization circuits, significantly reducing hardware architecture complexity while maintaining sufficient computational accuracy for the application.
Solution Approach 2:
By changing the numerical representation parameters from floating-point to quantized formats, the patent simplifies the hardware requirements. The quantized operations can be implemented with basic adders, shifters, and comparators, avoiding the need for complex floating-point processing units, thus reducing device complexity.
3Ease of operation
If CNNs are implemented with standard floating-point hardware, then ease of operation is maintained, but productivity and processing speed are limited
Solution Approach 1:
The patent changes the computational parameters to use quantized arithmetic that maps directly to efficient hardware operations. This allows the system to achieve high processing speeds through simple integer arithmetic and bitwise operations while maintaining compatibility with standard software development practices through appropriate API interfaces.
Data Source
AI summary
A sequence of convolutional operations of a quantized convolutional neural network (quantized CNN) is performed on an input video frame using quantized weights to generate feature maps. Respective batch normalizations are applied to the feature maps to obtain normalized feature maps. Applying a batch normalization to a feature map of the feature maps includes applying a linear function to the feature map where the linear function includes multiplying each feature of the feature map by a learned scaling factor. After applying the respective batch normalizations, the normalized feature maps are processed through additional layers of the quantized CNN to determine a probability that the input video frame is of low quality.