Image Encoder Pretraining for Low-Label Defect Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The labeling of defect samples from scanned images in product quality inspection is time-consuming and costly, making it difficult to train an image defect detection model effectively, especially in scenarios with few defect products.
Innovation Solution
An image encoder determination method that utilizes self-supervised training to optimize the image encoder by reconstructing image patch codes from scanned images of the same object under different lighting parameters, reducing the need for labeled defect samples and enabling training with a large number of normal samples.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional defect detection modeling is used requiring labeled defect samples, then the model can be trained to detect defects, but the labeling time and cost become excessively high
Solution Approach 1:
The system performs self-supervised learning where the model learns to encode and reconstruct images automatically without human labeling. The reconstruction task serves as the learning objective, enabling the model to develop feature extraction capabilities inherently useful for defect detection without requiring external labeling resources
Solution Approach 2:
The method pre-trains the image encoder using large amounts of normal product images before actual defect detection. This preliminary training phase builds robust feature representation that can be transferred to defect detection tasks, eliminating the need for time-consuming defect sample labeling
2Measurement precision
If traditional defect detection modeling is used requiring labeled defect samples, then the model can be trained to detect defects, but the labeling cost becomes excessively high
Solution Approach 1:
The system performs self-supervised learning where the model learns to encode and reconstruct images automatically without human labeling. The reconstruction task serves as the learning objective, enabling the model to develop feature extraction capabilities inherently useful for defect detection without requiring external labeling resources
Solution Approach 2:
The method pre-trains the image encoder using large amounts of normal product images before actual defect detection. This preliminary training phase builds robust feature representation that can be transferred to defect detection tasks, eliminating the need for time-consuming defect sample labeling
3Quantity of substance
If self-supervised training with normal samples is used, then the quantity of labeled defect samples is reduced, but the feature expression capability must be maintained or improved
Solution Approach 1:
The method pre-trains the image encoder using large amounts of normal product images before actual defect detection. This preliminary training phase builds robust feature representation that can be transferred to defect detection tasks, eliminating the need for time-consuming defect sample labeling
Solution Approach 2:
The reconstruction loss provides continuous feedback during training, guiding the encoder to learn meaningful feature representations. The model adjusts its parameters to minimize reconstruction error, which inherently preserves important visual features necessary for defect detection
Data Source
AI summary
This application discloses an image encoder determination method performed by a computer device. The method includes: inputting, for a first sample image and a second sample image of a first object under different lighting parameters, the first sample image into an image encoder in an initial reconstruction model for image encoding, and outputting first image patch codes respectively corresponding to a plurality of first image patches; inputting the plurality of first image patch codes into a reconstruction network in the initial reconstruction model, and performing code prediction on a plurality of second image patches in the second sample image to output a plurality of first predicted codes; and performing model training on the initial reconstruction model with reference to a plurality of second image patch codes obtained by inputting the plurality of second image patches into a pre-trained encoder and a loss function, to obtain a first reconstruction model.