Compared with wavelet analysis, the X-let multi-scale geometric analysis tool realizes a more efficient sparse representation of images, that is, image information will be more concentrated in a small number of coefficients in the transform domain, which is conducive to the separation of image signals and noises. At the same time, there are more directional subbands at the same scale, and the combination of multiple sub-generational coefficient features can identify noise more accurately. Because it is generally believed that the transformation coefficients with large magnitudes in all subbands correspond to strong edges of the image, and the transformation coefficients with large magnitudes in some direction subbands but small magnitudes in other direction subbands at the same scale correspond to For the weak edge of the image, the magnitudes in all subbands are small-value coefficients corresponding to noise, which provides a premise for denoising. Thermal infrared images have both strong noise and weak edge characteristics, and the use of multi-scale geometric analysis can naturally achieve image denoising and enhancement, so it has obvious advantages in image preprocessing.

 

The general steps of multi-scale geometric analysis denoising are as follows:

(1) Perform multi-scale geometric transformation on the infrared image to obtain multi-scale and multi-directional subband coefficients;

(2) Analyze the characteristics of the coefficients in different sub-bands and deal with them accordingly. For example, the strong edge coefficient should be kept, the weak edge coefficient should be enlarged (or maintained), and the noise figure should be zeroed

(3) Multi-scale geometric analysis and inverse transformation to obtain denoised images.