Dataset Generation System for Neural Network Fine-Tuning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In large chain stores, fine-tuning of deep neural network models for image analysis across numerous facilities is resource-intensive and costly due to the need for extensive data transfer and storage, especially when using extremely light models for edge-side devices.
Innovation Solution
A dataset generation system that classifies cameras into groups, collects and evaluates captured images based on user-defined criteria, and uses inference processing to determine suitability for neural network training, reducing the number of images required for model learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If fine tuning of the original learned DNN model is performed using captured images of all cameras in several 1000 stores, then inference accuracy is improved, but processing time increases and communication and storage costs increase
Solution Approach 1:
The patent segments the large-scale fine-tuning process into two phases: (1) a first phase performing fine-tuning on a small number of selected cameras to generate an intermediate model, and (2) a second phase performing fine-tuning on the remaining cameras using the intermediate model. This segmentation reduces the processing time and costs while maintaining inference accuracy by avoiding the need to process all cameras simultaneously in a single exhaustive fine-tuning operation.
2Measurement precision
If fine tuning of the original learned DNN model is performed using captured images of all cameras in several 1000 stores, then inference accuracy is improved, but communication cost and storage cost increase
Solution Approach 1:
The patent divides the fine-tuning process into stages, first processing a small subset of cameras to create an intermediate model, then using this intermediate model to process remaining cameras. This reduces the total volume of data that needs to be transferred and stored, thereby lowering communication costs (data transfer between edge devices and servers) and storage costs (server storage requirements), while still achieving accurate inference across all facilities.
3Device complexity
If an extremely light DNN model is used in edge-side devices, then device resource requirements are reduced, but fine tuning requires extensive data processing across numerous facilities
Solution Approach 1:
The patent performs preliminary fine-tuning on a small subset of cameras first to generate an intermediate model before proceeding to fine-tune the remaining cameras. This preliminary action creates a refined base model that serves as a better starting point for subsequent fine-tuning operations, thereby improving overall fine-tuning efficiency while maintaining the benefits of using extremely light DNN models in edge devices.
Data Source
AI summary
A dataset generation system has a camera classification circuitry configured to classify plural cameras into plural groups, an input device for a user to set a selecting criterion of a captured image, a first captured image collection circuitry configured to collect captured images which are captured by at least one camera in each group classified by the camera classification circuitry, and which meet the selecting criterion set by a user using the input device, an inference circuitry configured to perform inference processing on each of the captured images collected by the first captured image collection circuitry and a dataset evaluation circuitry configured to evaluate whether or not a dataset consisting of the captured images collected by the first captured image collection circuitry is suitable for a training dataset of a neural network model for a predetermined inference process, based on the result of the inference processing by the inference circuitry.