Neural Network Image Processing System with Segmented Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In factory automation, image processing devices using convolutional neural networks (CNNs) face challenges with calculation performance when inspecting objects, leading to insufficient determination times and inability to inspect in real time, especially when the measurement target is not a learned object, resulting in low function devices being unable to perform accurate inspections.
Innovation Solution
An image processing system is implemented with a low function device and a high function device, where the low function device applies a first neural network to captured images and outputs results to the high function device when accuracy is below a threshold, allowing the high function device to apply a second neural network, partially common with the first, to enhance determination accuracy and perform relearning for improved performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a low function device uses a neural network for image processing, then the device can perform automated inspection, but the calculation performance is insufficient leading to inability to finish determination within a certain period of time
Solution Approach 1:
The patent divides the neural network into multiple stages: a first stage neural network executed by the low function device for initial processing, and a second stage neural network executed by a high function device for final determination. This segmentation allows the low function device to perform preliminary filtering while complex calculations are offloaded to the high function device, resolving the contradiction between automated inspection capability and determination time.
Solution Approach 2:
The patent introduces an image management device as an intermediary that coordinates between the low function device and high function device. This intermediary manages the transfer of images and determination results, optimizing the workflow and ensuring efficient utilization of both low and high function devices, thereby improving overall inspection speed while maintaining acceptable determination times.
2Measurement precision
If a low function device constructs a network model with large calculation amount, then the model can achieve high accuracy, but the calculation performance becomes insufficient
Solution Approach 1:
The patent segments the neural network model into two parts: a first neural network with reduced calculation amount executed by the low function device, and a second neural network with larger calculation amount executed by the high function device. This segmentation enables high accuracy to be achieved through the second network while the low function device handles only the lighter initial processing, resolving the contradiction between accuracy and calculation capability.
Solution Approach 2:
The patent creates a simplified copy of the neural network (first neural network) that runs on the low function device for preliminary processing. This copy performs initial feature extraction and filtering, reducing the burden on the low function device while maintaining the capability to achieve high accuracy through the complete network on the high function device.
3Adaptability or versatility
If a low function device performs learning to optimize parameters, then the device can adapt to different inspection targets, but the calculation performance required for learning cannot be provided
Solution Approach 1:
The patent introduces a high function device as an intermediary that performs the computationally intensive learning and parameter optimization. The low function device focuses on executing the neural network for inspection, while the high function device handles model training and parameter tuning. This division enables adaptability through learning without requiring the low function device to have high learning calculation capability.
Solution Approach 2:
The patent creates a simplified version of the neural network model that can be deployed on the low function device after learning is performed elsewhere. This allows the system to maintain adaptability through optimized parameters while the low function device only needs to execute the pre-learned model, avoiding the need for high learning calculation capability at the edge device.
Data Source
AI summary
Provided is an image processing system which determines at least one object using a low function device and a high function device having higher calculation capacity than the low function device. The low function device includes a part configured to output a first determination result which identifies the objects by applying a first convolution neural network to captured images of the objects, and a part configured to output the captured image to the high function device when the first determination result meets a predetermined condition. The high function device includes a part configured to apply a second convolution neural network preliminarily learned from a sample which is at least partially common to the first neural network to the captured image and to output a second determination result which identifies the object.