Autoregression Image Anomaly Detection via Latent Space Memory
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image abnormity detection methods in computer vision face challenges with large data distribution and variance, particularly with deep autoencoders lacking effective solutions for these issues, and requiring clear supervision information that is difficult to obtain.
Innovation Solution
An autoregression image abnormity detection method enhancing a latent space based on memory, which includes constructing a network structure with an autoencoder, autoregression module, and memory module to model data distribution without prior distribution settings, allowing for effective abnormal image identification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If deep autoencoder is used for abnormity detection, then automatic characteristic learning is achieved, but large data distribution and large data variance problems occur
Solution Approach 1:
The patent segments the latent space into multiple subspaces by introducing subspace identification modules. Each subspace is trained independently to handle specific portions of the data distribution, which prevents any single autoencoder from being overwhelmed by large data variance and distribution complexity. This segmentation allows the system to maintain automatic learning while managing data distribution stability.
Solution Approach 2:
The patent introduces a new dimension of subspace management alongside the traditional latent space. By organizing latent representations into multiple subspaces with different dimensional characteristics, the system can better capture and manage data distribution variations without compromising the automatic learning capability of the deep autoencoder.
2Extent of automation
If deep autoencoder is used for abnormity detection, then automatic characteristic learning is achieved, but abnormal image reconstruction occurs
Solution Approach 1:
The patent implements feedback mechanisms through reconstruction loss calculation and subspace validation. The system continuously monitors reconstructed images and uses this feedback to identify and correct abnormal reconstructions. When abnormal images are detected in the latent space, the system adjusts the corresponding subspace representations to prevent future abnormal reconstructions.
Solution Approach 2:
The patent converts the harmful effect of abnormal image reconstruction into a beneficial detection signal. By allowing the autoencoder to attempt to reconstruct abnormal images and then measuring the reconstruction error, the system creates a detection mechanism that identifies abnormal samples. The harmful reconstruction attempt becomes the basis for useful abnormality detection.
3Stability of the object's composition
If prior distribution settings are used, then data distribution can be controlled, but data distribution is damaged
Solution Approach 1:
The patent introduces dynamic subspace adaptation where the latent space subspaces are continuously adjusted during training to match the actual data distribution. Instead of using fixed prior distribution settings, the system dynamically learns and adapts the subspace structures from the data itself, maintaining data distribution integrity while still providing control over the representation space.
Solution Approach 2:
The patent changes the parameters of the latent space representation by introducing multiple subspaces with different dimensional and statistical characteristics. This allows the system to control and organize data distribution without imposing damaging prior distribution settings, as the subspace parameters are learned from and adapt to the actual data characteristics.
Data Source
AI summary
The present application discloses an autoregression image abnormity detection method of enhancing a latent space based on memory, which belongs to the field of abnormity detection in computer vision. The present application comprises: selecting a training data set; constructing a network structure of an autoregression model of enhancing a latent space based on memory; preprocessing the training data set; initializing the autoregression model of enhancing a latent space based on memory; training the autoregression model of enhancing a latent space based on memory; verifying the model on the selected data set, and using the trained model to judge whether the input image is an abnormal image. In the present application, a prior distribution is not needed to be set such that the distribution of the data itself will not be destroyed, and it can prevent the model from reconstructing abnormal images, and ultimately can better judge abnormal images.