Variational Autoencoder Image Augmentation for Gas Leakage Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image augmentation methods using variational autoencoders face challenges with low reconstruction accuracy due to the size of input images, resulting in blurred reconstructed images.
Innovation Solution
The method constructs a variational learner and discriminator based on a fully convolutional neural network, allowing for the extraction of features from images of any size, and adjusts the variational autoencoder model to improve reconstruction accuracy by training with a gas leakage image until a preset threshold is met, ensuring clear augmented images are generated.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If a variational autoencoder is used for image augmentation, then the image can be reconstructed, but the reconstructed image is blurred and has low reconstruction accuracy
Solution Approach 1:
The patent segments the image processing task into multiple stages: original image processing, augmented image generation, and reconstructed image production. By dividing the reconstruction process and applying different processing strategies at each stage, the patent maintains image clarity while achieving accurate reconstruction, resolving the contradiction between reconstruction accuracy and image clarity.
Solution Approach 2:
The patent applies different quality standards and processing methods to different parts of the image pipeline. The original image is preserved with full quality, the augmented image is generated with enhanced features, and the reconstructed image is processed to maintain clarity. This local quality approach ensures that each stage optimizes for its specific purpose without compromising overall reconstruction accuracy and image clarity.
2Ease of manufacture
If the size of the input image is considered in variational autoencoder, then the model can be trained, but the reconstructed image quality deteriorates
Solution Approach 1:
The patent changes key parameters in the variational autoencoder model, including modifying the encoder and decoder structures, adjusting the loss function weights, and optimizing the training parameters. These parameter changes enable the model to be trained effectively on images of various sizes while maintaining high reconstruction accuracy and preventing quality deterioration.
3Ease of operation
If traditional image augmentation methods are used, then the process is simple, but the reconstruction accuracy is low and images are blurred
Solution Approach 1:
The patent introduces an intermediary augmented image generation stage between the original image and the reconstructed image. This intermediary stage uses a trained variational autoencoder to generate augmented images that preserve important features while allowing for flexible transformation. This mediator enables simple operation through automated processing while achieving high reconstruction accuracy that traditional methods cannot attain.
Data Source
AI summary
An image augmentation method applied to an electronic device is provided. The method includes constructing a variational learner and a discriminator based on a fully convolutional neural network. A target image is obtained by inputting a gas leakage image into the variational learner. A variational autoencoder model is obtained by training the variational learner based on a discrimination result of the discriminator on the target image. A reconstruction accuracy rate is calculated based on a test image, an augmented model is obtained by adjusting the variational autoencoder model based on the gas leakage image, in response that the reconstruction accuracy rate being less than a preset threshold; and an augmented image is obtained by inputting the image to be augmented into the augmented model.