Image Inspection Segmentation With A Fixed Encoder
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
U-Net requires a large computational cost for learning, making it difficult to implement on equipment with relatively low processing capability, such as CPUs, which hinders its use in production sites for image inspection.
Innovation Solution
An image inspection apparatus using a machine learning model with a fixed encoder part and adjustable decoder and connection parts, where parameters are updated based on training data, reducing computational requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If U-Net is used for image inspection, then segmentation accuracy is improved, but computational cost increases
Solution Approach 1:
The patent divides the U-Net model into two separate parts: an encoder part and a decoder part. The encoder part is trained offline using pre-collected images, while the decoder part is trained online using images captured during production. This segmentation allows the computationally intensive encoder training to be performed in advance, reducing the computational burden during production while maintaining segmentation accuracy.
Solution Approach 2:
The patent performs preliminary training of the encoder part using pre-collected images before actual production inspection. This preliminary action transfers the heavy computational workload to the offline stage, enabling the model to achieve accurate segmentation without requiring high computational resources during production runtime.
2Measurement precision
If U-Net is used for image inspection, then segmentation accuracy is improved, but learning time increases
Solution Approach 1:
The patent segments the training process into offline encoder training and online decoder training. This allows the time-consuming encoder training to be performed in advance when time is not constrained, while the quick decoder training occurs during production setup. The result is accurate segmentation with minimal impact on production timelines.
Solution Approach 2:
The encoder part is trained in advance using pre-collected images before production begins. This preliminary training action eliminates the need for lengthy training during production setup, reducing the effective learning time that impacts production while maintaining high segmentation accuracy.
3Adaptability or versatility
If full model training is performed on-site, then model adaptability is improved, but device complexity requirements increase
Solution Approach 1:
The patent divides the model into encoder and decoder parts with different training schedules. The encoder, which requires complex computations, is trained offline. The decoder, which has lower computational requirements, is trained on-site using production images. This segmentation enables model adaptability to production conditions without requiring high processing capability equipment at the production site.
Solution Approach 2:
The patent changes the training parameters and timing for different model parts. The encoder is trained with comprehensive pre-collected data offline, while the decoder is trained with production images online. This parameter change in training strategy allows the model to adapt to production conditions using equipment with limited processing capability.
Data Source
AI summary
An image inspection apparatus includes a control unit configured to execute a model which is a segmentation model. The model includes an encoder part configured to extract a first feature from the inspection image data, a connection part configured to receive a second feature different from the first feature from at least one of layers in the encoder part, convert the second feature into a third feature, and supply the third feature, and a decoder part configured to upsample the first feature using the third feature. The control unit updates the parameters of the connection part and the parameters of the decoder part when executing machine learning of the model based on the training image data.