Imaging Pixel Exposure Control for CNN Convolution Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Convolution processing in CNNs requires a large processing load due to the need for extensive calculations across all imaging pixels, leading to significant computational demands.
Innovation Solution
An information processing device that sets exposure times for imaging pixels based on convolution coefficients and transfers signal charges to a floating diffusion for analog convolution processing, reducing the need for extensive arithmetic operations in the CNN's feature extracting layer.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If convolution processing is performed using a plurality of surrounding pixels for each target pixel in a CNN, then feature extraction accuracy is improved, but processing load increases significantly
Solution Approach 1:
The patent replaces digital arithmetic operations with optical processing. Specifically, it uses optical convolution processing where light modulation elements (such as spatial light modulators or microlens arrays) physically perform the convolution operation through optical diffraction and interference patterns, eliminating the need for extensive digital calculations while maintaining feature extraction accuracy
Solution Approach 2:
The patent changes the computational domain from digital arithmetic to optical parameters. By encoding image data and convolution kernels into optical fields (amplitude, phase, wavelength), the system performs convolution through optical transformations rather than digital multiplication and addition, fundamentally altering the processing mechanism to reduce computational load
2Reliability
If convolution processing is performed on all necessary regions in a CNN, then comprehensive feature extraction is achieved, but processing time increases
Solution Approach 1:
The patent performs convolution processing for multiple layers simultaneously in a single optical transformation step. By encoding multiple convolution kernels and their corresponding layer operations into the optical system's configuration (through multiplexing techniques), the system completes what would traditionally require sequential processing across all regions in parallel, significantly reducing processing time while maintaining completeness
Solution Approach 2:
The patent implements continuous optical processing where the optical convolution operation processes the entire image field continuously without discrete computational steps. The optical system maintains continuous wavefront propagation and transformation throughout the processing region, eliminating the step-by-step computational interruptions inherent in digital processing and enabling comprehensive feature extraction in real-time
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
This approach significantly reduces the processing load of convolution processing by performing it in the imaging unit, allowing the subsequent layers to focus on pooling and classification without the need for initial convolution processing, thereby enhancing efficiency.
Implementation Method 1
an imaging unit including a plurality of imaging pixels arrayed two-dimensionally
Data Source
AI summary
Provided are an information processing device, an information processing method, and an information processing program capable of reducing a processing load of convolution processing in a convolutional neural network (CNN). An information processing device (1) according to the present disclosure includes a setting unit (51) and a control unit (52). The setting unit (51) sets exposure time of each of imaging pixels in an imaging unit (2), which includes a plurality of imaging pixels arrayed two-dimensionally, to exposure time corresponding to a convolution coefficient of a first layer of a CNN. The control unit (52) causes transfer of signal charges from imaging pixels, which have been exposed, to a floating diffusion (FD), thereby performing convolution processing.