Multi-Scale Autoencoder Reconstruction for Product Defect Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The size of feature maps generated by autoencoders can be too large or too small, leading to a low degree of fineness in reconstructed images, making them dissimilar to the original input images.
Innovation Solution
A multi-scale autoencoder generation method that constructs latent spaces of varying scales, iteratively calculates ratios to control feature map sizes, trains autoencoders using non-defective images, and calculates similarity values to reconstruct and detect defects in product images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If a single-scale autoencoder is used, then the model structure is simple, but the reconstructed image has low fineness and is dissimilar to the original image
Solution Approach 1:
The patent divides the single autoencoder into multiple autoencoders with different scales (first autoencoder, second autoencoder, etc.). Each autoencoder processes feature maps at different resolution levels, segmenting the reconstruction task into multiple specialized sub-tasks that collectively achieve high-fidelity reconstruction.
Solution Approach 2:
The patent introduces a multi-scale dimension by creating autoencoders with different feature map sizes (e.g., 64x64, 32x32, 16x16). This dimensional variation allows the system to capture both global and local image characteristics, improving reconstruction quality without requiring excessive complexity in any single model.
2Measurement precision
If feature map size is increased, then more image details are captured, but computational complexity and memory usage increase
Solution Approach 1:
The patent employs multiple autoencoders that each process only specific scale levels rather than one autoencoder handling all scales. This partial action approach distributes the computational burden, where each autoencoder operates on feature maps of appropriate size for its specific task, avoiding the excessive computational cost of a single large-scale autoencoder.
Solution Approach 2:
The patent changes the scale parameter (feature map size) across different autoencoders. By varying this parameter systematically (creating a hierarchy of scales), the system optimizes the balance between detail capture and computational efficiency for each autoencoder, rather than using a uniform scale that would compromise either precision or efficiency.
Data Source
AI summary
A multi-scale autoencoder generation method applied to an electronic device is provided. The method includes acquire product images and acquire an annotation of each product image. Latent spaces of a plurality of scales are constructed. Autoencoders are obtained according to the latent spaces and an image size of the product image. Learners are obtained by training each autoencoder based on non-defective images. Reconstructed images are obtained by inputting the product images into the learners. Detection results are obtained by detecting whether each product image has defects according to the reconstructed images. Similar images for each learner are determined based on a comparison result between each detection result and a corresponding annotation result. Once a correct rate of each learner is obtained according to the similar images, a learner from the plurality of learner is determined as a multi-scale autoencoder according to the correct rate of each learner.