Group-Specific Neural Network Models for Diverse Camera Conditions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing edge-side devices with insufficient computer resources struggle to perform highly accurate object detection and recognition using extremely light learned DNN models, especially when deployed across numerous facilities, due to diverse image conditions and the impracticality of fine-tuning or transfer learning for each camera.
Innovation Solution
A group-specific model generation system that collects and groups images based on features, classifies cameras, and performs fine-tuning or transfer learning on neural networks for each group, generating specialized models suitable for the specific conditions of each group, even with limited resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If an extremely light learned DNN model is implemented on edge-side devices, then the processing load is reduced and the device resources are sufficient, but the object detection and recognition accuracy deteriorates
Solution Approach 1:
The patent segments the large number of facilities into multiple groups based on image features (layout, light conditions, interior decoration). Instead of treating all facilities uniformly, each group receives a specialized learned DNN model trained on representative images from that group, thereby improving accuracy while keeping individual models lightweight.
Solution Approach 2:
The patent changes the parameter of model specialization by training separate learned DNN models for different facility groups. Each group-specific model is trained using image data characteristic of that group's environment, allowing the model parameters to adapt to specific facility conditions while maintaining light weight for edge device deployment.
2Measurement precision
If fine-tuning or transfer learning is performed for each camera to ensure accuracy, then the object detection and recognition accuracy is improved, but the time and resources required become enormous and impractical
Solution Approach 1:
The patent merges the fine-tuning process by training a single learned DNN model using image data from multiple facilities within the same group. Instead of performing separate fine-tuning for each camera, the model learns from aggregated group data, significantly reducing the total time and computational resources required while still adapting to group-specific characteristics.
Solution Approach 2:
The patent applies partial fine-tuning at the group level rather than complete fine-tuning for each individual camera. This partial action approach trains models on representative group data without requiring exhaustive per-camera fine-tuning, achieving sufficient accuracy while keeping the process practical and resource-efficient.
3Quantity of substance
If fine-tuning or transfer learning is performed using captured images from all cameras in several thousand stores, then more data is available for training, but the diversity of image conditions prevents sufficient learning of the light DNN model
Solution Approach 1:
The patent segments the diverse image data from all cameras into distinct groups based on image features such as layout, light conditions, and interior decoration. By training separate learned DNN models for each group rather than using all data uniformly, the system ensures that each model learns from homogeneous data characteristics, improving learning effectiveness despite the large volume of diverse training data.
Data Source
AI summary
A group-specific model generation system includes: a captured image collection circuitry configured to collect a captured image from each of cameras; an image feature extraction circuitry configured to extract a feature from each of the captured images; an image clustering circuitry configured to group the captured image based on the feature of each of the captured images; a camera classification circuitry configured to classify cameras into groups, based on a result of the grouping of the captured images; and a group-specific model generation circuitry configured to generate a group-specific learned neural network model suitable for captured images taken by the cameras in the each of the groups.