Infrared Image Super-Resolution Method and Apparatus Using Nonsubsampled Contourlet Transform

Infrared Super-Resolution Using Nonsubsampled Contourlet Transform and Dual Sparse Dictionaries

This technology is a super-resolution method and apparatus that reconstructs low-resolution infrared images into high-resolution by learning raw sparse dictionary pairs through nonsubsampled contourlet transform of training images, additionally learning residual sparse dictionary pairs from the residuals of the initial reconstruction, and referencing both dictionary pairs.

Infrared thermal cameras often suffer from low resolution due to hardware limitations, frequently leading to failed subsequent analysis. Existing learning-based super-resolution techniques have limitations, as learned dictionaries are unstructured, simple derivative filters fail to sufficiently represent image features, and a single dictionary pair is often insufficient for satisfactory reconstruction.

This technology structures sparse dictionaries by creating coefficient blocks using nonsubsampled contourlet transform, which enables directional multi-scale representation, and compensates for high-frequency components missed by the raw dictionary pair with a residual dictionary pair. During testing, the final high-resolution infrared image is obtained by sequentially integrating the estimated high-frequency map and estimated residual image into a bicubic upscaled image. Applicable to security surveillance, defense reconnaissance, and biomedical thermal imaging, it allows for precise identification of targets and heat distribution without expensive high-resolution infrared sensors.

‍

‍

Key Features:

• Step of creating training low-resolution coefficient blocks by performing nonsubsampled contourlet transform on training LR IR images and forming training high-resolution coefficient blocks from training high-frequency maps
• Raw sparse dictionary generation step of creating raw low-resolution and high-resolution sparse dictionaries using the training coefficient blocks
• Step of forming residual high-resolution and low-resolution sparse dictionaries by creating training residual IR images from training HR IR images and estimated training HR IR images
• Step of creating an estimated raw HR IR image through sparse representation of a test LR IR image and forming the final HR IR image by integrating it with the estimated residual IR image

‍