Facies-Guided Seismic Super-Resolution With Pixel Attention
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing seismic image super-resolution techniques face challenges in achieving high-resolution images efficiently, as they require high computational and financial costs, and existing deep learning models struggle with training stability and scalability, especially when upsampling factors increase.
Innovation Solution
A self-calibrated convolution neural network with a pixel attention block and facies classification loss is employed to guide upsampling, using a combination of Depthwise Separable Convolution layers and a facies classifier to enhance seismic image resolution while incorporating facies information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If conventional super-resolution techniques (interpolation, contour features, statistical image priors) are used, then the method is simple to implement, but the mapping definition between LR and HR images is unclear and the process is inefficient for establishing complex high-dimensional maps
Solution Approach 1:
The patent replaces conventional mechanical image processing methods (interpolation, contour features) with a deep learning-based neural network system. The neural network learns complex non-linear mappings between low-resolution and high-resolution seismic images through training on paired data, eliminating the need for manual definition of mapping rules while achieving superior reconstruction accuracy.
Solution Approach 2:
The patent transforms the approach by changing from fixed geometric transformation parameters to learned data-driven parameters. The neural network automatically adapts its internal parameters during training to capture the complex relationship between LR and HR images, allowing flexible handling of various upsampling factors and seismic characteristics without requiring explicit mapping definitions.
2Measurement precision
If SRCNN size is increased to improve SR accuracy, then more parameters are used, but real-time execution becomes difficult and training settings become hard to determine
Solution Approach 1:
The patent optimizes the neural network architecture by adjusting key parameters such as number of layers, filter sizes, and network depth to achieve an optimal balance between accuracy and computational efficiency. The system determines appropriate training settings (learning rate, weight initialization, weight decay) through systematic parameter tuning, enabling real-time execution while maintaining high SR accuracy for seismic image processing.
3Device complexity
If light SR model with recursive depth-wise separable convolution is used, then model complexity is reduced, but the number of multiply-accumulate operations increases due to recursive operation
Solution Approach 1:
The patent extracts and removes the problematic recursive operation from the light SR model while retaining its depth-wise separable convolution structure. By replacing the recursive block with a non-recursive alternative that maintains computational efficiency, the system reduces the number of multiply-accumulate operations while preserving the model's ability to handle super-resolution tasks effectively.
4Adaptability or versatility
If GAN-based model with four convolutional layers is used, then adaptive bandwidth extension is achieved, but training becomes unstable and unsupervised learning makes the model hard to train and generate output
Solution Approach 1:
The patent introduces a feedback mechanism through the use of discriminator networks and loss functions that provide guidance during training. The supervised learning approach with paired LR-HR seismic image data enables stable training by providing clear target values for optimization, while the adaptive bandwidth extension capability is maintained through the network's ability to learn frequency-domain transformations during training.
Data Source
AI summary
A facies classification apparatus and method of training the apparatus is provided for seismic image super-resolution. The apparatus includes an input for receiving a low resolution seismic images extracted from a seismic volume and a feature extraction section to extract features from the low resolution images. A non-linear feature mapping section generates feature maps using a self-calibrated block with pixel attention having a plurality of Depthwise Separable Convolution (DSC) layers. A late upsampling section combines at least one DSC layer that upsamples the feature maps to a predetermined dimension. An output provides approximate upsampled super-resolution seismic images that correspond to the low resolution seismic images, with a desired scale of at least two times the low resolution seismic image. A pretrained facies classifier classifies facies in the seismic volume based on the approximate upsampled super-resolution seismic images to obtain class labels.