Deconvolution
Powerful deblurring and sharpening.
Deconvolution is a very powerful image processing tool that can sharpen and restore detail. A lot of progress in image sharpening and deblurring has been done using AI models. Many of these models use generative techniques. The deconvolution tools in Astra Image provide repeatable results with no extra information being generated.
One of the key points for successful deconvolution is to accurately define the blur kernel (also known as the point spread function, or PSF). The blur kernel defines the imperfection that caused a problem with the image. For example, an image usually has a Gaussian blur kernel due to lens imperfections and natural air movements. Therefore, if you define a small Gaussian blur kernel, the image can be sharpened as the effects of the lens and air movements can be minimized. On the other hand, an image taken as the camera is moving will have a blur kernel that looks like the motion of the camera during the exposure. When running deconvolution, please remember that even a small adjustment to the blur kernel parameters can have a big impact on the final result.
Deconvolution Type Five different deconvolution methods are available in this version of Astra Image:
-Van Cittert Good for sharpening. Not good when significant noise is present, but the strength can be adjusted to mitigate the effect of noise to some degree.
-Jansen Van Cittert Good for sharpening. Not as effective when significant noise is present.
-Landweber Good balance between sharpening and noise suppression.
-Lucy Richardson Good for sharpening and restoration, even when noise is present.
-Maximum Entropy Good for restoration even when significant noise is present.
Target Data Deconvolution can be performed on either the luminance channel, or on each red, green and blue channel separately. If image sharpening is the goal, the luminance channel can be used. As only one channel is being processed, this will be the fastest method. For image restoration, such as removing motion blur, better results are usually obtained by processing the red, green and blue channels.
Processing Backend To get the fastest results, the deconvolution process can be run on a GPU. To do this, either CUDA (for nVidia GPUs) or OpenCL (for ATI, Intel GPUs) must be installed on your system. If a GPU is not available, a highly-optimized CPU version will be used.
Iterations Deconvolution is an iterative process that ultimately converges at a solution. Once the algorithms approach the solution, the effect of deconvolution gets less and less. Therefore, the number of iterations needs to be enough to reach a solution. A good starting point is between 5 and 15 iterations.
Strength When using Van Cittert or Lucy Richardson methods, you can also set the strength of the deconvolution. A higher number will produce a more pronounced result, but will also result in higher noise amplification. When dealing with noisy images, reducing the strength can produce better results.
Convert to Linear Pixel Values Some images use the sRGB color space. This can result in a less than optimal result, so this option will convert from sRGB into linear RGB for processing and then back into sRGB when the processing is finished.
Blur Kernel There are two types of blur kernel: parametric and motion blur.
-Parametric This will generate a 2 dimensional curve that is usually used for sharpening or restoring out-of-focus images.
-Motion Blur This will generate a straight line that can be used to restore an image when there is linear motion blur present.
Size When using a parametric blur kernel, this sets the size of the curve (sigma). When using a motion blur blur kernel, this sets the length of the line.
Aspect Ratio On many occasions, the blur kernel will not be a symmetric curve. For example, in an astronomical image, the stars may be slightly elogated due to imperfect tracking. The aspect ratio of the parametric blur kernel allows you to make elongated blur kernels.
Angle of Rotation When using a parametric blur kernel with an aspect ratio of less than 1.0, the elongated blur kernel can be rotated. When using a motion blur blur kernel, you can rotate the line to match the direction of the blur.
Kurtosis This sets the steepness of the sides of the curve. When the kurtosis is set to 2.0, the result will be a Gaussian curve. When set to less than 2.0, the curve will get flatter. When set to greater than 2.0, the curve will get steeper. Higher settings can sometimes be useful for correcting images that are out of focus.
Preview Process the image without making an permanent changes.
Last updated
