Image Recognition System Using Reference Image Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image recognition systems using deep learning require a large computational load when processing series image sequences, such as time-series continuous or multi-spectral images, due to the need for extensive calculations across all images in the sequence.
Innovation Solution
An image recognition system that selects a reference image and a proximity image from a series image sequence, performing inference processing using convolution and activation functions on the reference image and generating results for the proximity image by approximating the processing results from the reference image and the difference image between the two, reducing redundant calculations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning methods are used for image recognition on series image sequences, then recognition accuracy is improved, but computational load increases extremely
Solution Approach 1:
The system performs inference processing on a reference image in advance from the series image sequence. The results of this preliminary inference are then reused when processing proximity images, avoiding redundant calculations and significantly reducing the overall computational load while maintaining recognition accuracy
Solution Approach 2:
The inference processing results obtained from the reference image are made universally applicable to multiple proximity images. By selecting multiple proximity images that are close to the reference image in time or feature space, the system reuses the same inference results across multiple processing tasks, reducing redundant computation
2Productivity
If student model is applied to all images in series image sequence according to distillation learning, then convergence speed is improved, but redundant calculations increase
Solution Approach 1:
The system performs inference processing on a reference image in advance from the series image sequence. The results of this preliminary inference are then reused when processing proximity images, avoiding redundant calculations and significantly reducing the overall computational load while maintaining recognition accuracy
Solution Approach 2:
Instead of discarding the inference results from the reference image, the system recovers and reuses them for processing proximity images. This prevents redundant calculations by utilizing previously computed results, thereby reducing energy consumption while maintaining fast convergence
Data Source
AI summary
A system includes: a unit that input a series image sequence; a unit that selects a reference image; a unit that selects a proximity image; and an inference unit that recognizes the reference image and the proximity image by performing inference processing, including convolution processing and activation function processing, on the reference image and the proximity image. The inference unit generates results of performing the convolution processing and the activation function processing on the proximity image from the results of the convolution processing and the activation function processing performed on the reference image, and the results of the product of the results of the convolution processing performed on a difference image, which is an image of the difference between the reference image and the proximity image, and a derivative value of the results of the convolution processing and the activation function processing performed on the reference image.