Deep Learning Vision Interface for Low-Power On-Board Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine vision systems face challenges in efficiently utilizing deep learning processes due to high computational demands, requiring advanced training procedures and interfaces that are often complex and require specialized expertise.
Innovation Solution
A vision system with an on-board processor of low to modest power, combined with a graphical user interface (GUI) that allows non-expert users to train and refine deep learning tools using confidence scoring and incremental learning, enabling training on a remote processor and runtime operation on the on-board processor.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning processes are applied to machine vision systems, then accuracy and adaptability are improved, but computational demands and device complexity increase
Solution Approach 1:
The system divides the deep learning workflow into two distinct phases: training phase performed on remote servers with high computational power, and inference phase executed on the on-board processor with limited power. This segmentation allows the machine vision system to achieve high accuracy through deep learning while keeping runtime computational demands low.
Solution Approach 2:
The system performs the computationally intensive training of deep learning models in advance on remote servers before deploying them to the on-board processor. By completing the heavy computational work beforehand, the system achieves high accuracy without requiring sustained high computational power during actual vision tasks.
2Measurement precision
If advanced training procedures are used to improve deep learning model performance, then model accuracy is improved, but ease of operation deteriorates due to complexity
Solution Approach 1:
The system introduces a remote server as an intermediary that handles the complex training procedures. Non-expert users can simply provide data and receive trained models without needing to understand or configure the sophisticated training processes, thus maintaining ease of operation while achieving high model accuracy.
Solution Approach 2:
The system enables automated model training and refinement where the deep learning framework automatically improves models through incremental learning on the remote server. This self-service capability eliminates the need for manual intervention in complex training procedures while continuously improving model accuracy.
3Measurement precision
If specialized expertise is required for training procedures, then deep learning model performance is improved, but ease of operation deteriorates
Solution Approach 1:
The remote server acts as an intermediary that encapsulates specialized deep learning expertise within its automated training framework. Non-expert users interact with a simplified interface while the remote server handles all specialized training operations, thus achieving high model performance without requiring users to possess specialized expertise.
4Measurement precision
If high computational resources are allocated for training, then deep learning tool performance is improved, but loss of time in deployment increases
Solution Approach 1:
The system performs all computationally intensive training operations in advance on remote servers before deployment. By completing the heavy computational work beforehand, the system achieves high deep learning tool performance while minimizing deployment time, as the on-board processor only needs to execute the already-trained model.
Solution Approach 2:
The system separates training and deployment into distinct phases executed at different locations and times. Training is performed in advance on remote servers, while deployment involves only loading and executing the pre-trained model on the on-board processor, thus achieving high performance without extending deployment time.
Data Source
AI summary
This invention overcomes disadvantages of the prior art by providing a vision system and method of use, and graphical user interface (GUI), which employs a camera assembly having an on-board processor of low to modest processing power. At least one vision system tool analyzes image data, and generates results therefrom, based upon a deep learning process. A training process provides training image data to a processor remote from the on-board processor to cause generation of the vision system tool therefrom, and provides a stored version of the vision system tool for runtime operation on the on-board processor. The GUI allows manipulation of thresholds applicable to the vision system tool and refinement of training of the vision system tool by the training process. A scoring process allows unlabeled images from a set of acquired and/or stored images to be selected automatically for labelling as training images using a computed confidence score.