Multiband Super-Resolution While Preserving Original Color
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current super-resolution algorithms for geospatial imagery, particularly satellite images, fail to maintain original color accuracy when applying super-resolution to individual bands, leading to increased false positives and reduced precision, and are ineffective for noisy images.
Innovation Solution
A deep learning-based approach using a very deep convolutional neural network (CNN) to apply super-resolution to individual bands of multispectral and hyperspectral images, minimizing pixel value modifications while maintaining original color, and incorporating a de-noising method to handle noisy pixels.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If current super-resolution algorithms are applied to combined color bands, then resolution is improved, but color accuracy deteriorates
Solution Approach 1:
The patent divides the image processing into separate channels by applying super-resolution algorithms to individual bands (red, green, blue) rather than processing combined color bands as a single unit. This segmentation allows each band to be enhanced independently, preserving color accuracy while improving resolution. The method processes R, G, and B bands separately through the super-resolution model and then combines them, avoiding the color distortion that occurs when merging bands before processing.
2Measurement precision
If super-resolution algorithms are applied to multispectral images with more than three bands, then resolution is improved, but algorithm effectiveness deteriorates
Solution Approach 1:
The patent extends the super-resolution algorithm to handle multispectral images with any number of bands (three or more), making the algorithm universally applicable. The method processes each spectral band independently through the same super-resolution model, allowing it to effectively handle RGB images, multispectral images with 4-10 bands, and hyperspectral images with 100+ bands. This universal approach maintains algorithm effectiveness across different image types while improving resolution.
3Measurement precision
If background processing algorithms are used to improve resolution, then resolution is improved, but false positive rate increases
Solution Approach 1:
By segmenting the processing into individual band handling rather than combined band processing, the patent reduces false positives. The method applies super-resolution to each spectral band independently, preserving the distinctive color signatures of objects. This prevents the color distortion and information loss that occur when bands are merged, thereby reducing false positive identifications while maintaining improved resolution.
Data Source
AI summary
System and methods for generating super resolution images from geospatial images having any number of bands. A super resolution model, which uses deep convolution neural networks (DCNNs), is trained using individual image bands, a large crop size or tile size of 512×512 pixels, and a de-noise algorithm. Applying one or more algorithms to maintain the original color of the image bands improves the quality metrics of the super resolution images as measured by PSNR (peak signal-to-noise ratio) and SSIM (structural similarity index measure) of super resolution images. Further applying one or more algorithms to remove border effects introduced during the disclosed process reduces and/or eliminates seam lines between tiles and enhances the overall accuracy of the super resolution images.