have a distorted (blurred image) and I perform several deconvolution algorithms on it to sharpen it (Richardson lucy, regularized inverse filter, tikhonov miller). And i want to measure each performance by calculating the MSE (mean square error). However I don’t have a reference or an ideal image where I can compare my result. is there a way to measure quality of sharpening when a reference image is not available ?

  • $\begingroup$ i think a general metric will not be very meaningful, but depending on content type, you could formulate some more relevant criteria $\endgroup$
    – tobalt
    Jul 29, 2023 at 8:49
  • $\begingroup$ thank you for your reply, can you please elaborate more on your suggestion. the content of my images are remote sensing images that is panchromatic $\endgroup$
    – gin
    Jul 29, 2023 at 13:58

1 Answer 1


To calculate MSE and other objective measures you obviously need a reference image. However, in the absence of a reference image, you can consider using some alternative approaches to assess the quality of image sharpening. Here are a few possibilities:

  1. Edge Preservation: Evaluate how well the sharpening algorithms preserve the edges in the image. Blurred images tend to have less defined edges, so a good sharpening algorithm should enhance edge details without introducing excessive artifacts or noise. You can use edge detection algorithms (e.g., Canny edge detector) on both the original and sharpened images and compare the results.

  2. Structural Similarity Index (SSIM): SSIM is a widely used metric for measuring the similarity between two images. It takes into account the luminance, contrast, and structural information. Although SSIM still requires a reference image, you can compare the SSIM values obtained from different deconvolution algorithms to determine which one produces results that are most similar to each other.

  3. Peak Signal-to-Noise Ratio (PSNR): While PSNR traditionally requires a reference image, you can calculate it by using the original blurred image as a reference. PSNR measures the ratio between the maximum possible power of a signal and the power of the noise. Although it's not a perfect measure for assessing image quality, it can provide some indication of the amount of noise introduced during the sharpening process.

Remember that these alternative approaches have limitations and may not provide the same level of accuracy as comparing against a ground truth image. However, they can still offer some insights into the performance of different deconvolution algorithms.

However, what I would personally do in this situation is to use a high-pass filter to measure the amount of high-frequency content in the images.

Here's a possible approach:

  1. Apply a high-pass filter to both the original blurred image and the sharpened image. A high-pass filter attenuates the low-frequency components of an image, allowing the high-frequency details to pass through.

  2. Calculate a metric that quantifies the amount of high-frequency content in the filtered images. One common and useful metric is the energy or variance of the high-frequency components. This can be computed by squaring the pixel values of the high-pass filtered images and calculating the mean or sum of the squared values.

  3. Compare the metric values between the original blurred image and the sharpened image. A successful sharpening algorithm should enhance the high-frequency content, resulting in a higher metric value for the sharpened image compared to the original blurred image.

While measuring the high-frequency content can provide valuable information about the effectiveness of image sharpening, keep in mind that even this is not a comprehensive measure of overall image quality. Other factors, such as noise, artifacts, and preservation of low-frequency details, also contribute to the perceived quality of the image. Therefore, it is beneficial to combine this metric with your own visual inspection (if possible) and other metrics mentioned earlier, to obtain a more comprehensive evaluation of the sharpening algorithms. You can give weights to each of these metrics and based on that you can form your overall evaluation!

  • 2
    $\begingroup$ “measure the amount of high-frequency content” deconvolution will increase this, sure. But algorithms can add way too much of it as well, and in the wrong places. This measure wouldn’t distinguish the “good” high frequencies from the “bad” ones. $\endgroup$ Jul 29, 2023 at 17:17
  • $\begingroup$ If you use high frequency content in your image, then only as something to minimize because deconv. can add anomalously much of it. The regularization tries to prevent it. $\endgroup$
    – tobalt
    Jul 30, 2023 at 5:17
  • $\begingroup$ @CrisLuengo I agree. Which is why it's best to combine it with other metrics and also to use one's own visual inspection to assist with the final evaluation. $\endgroup$ Jul 30, 2023 at 11:21

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.