Project

Question 1:

• To compare the PSNR values of frequency domain filtered images and spatial domain filtered images, we need to apply both filtering techniques on the same set of images using the same parameters. The PSNR values can then be calculated and compared to determine which technique gives better results.
• Regarding the statement that frequency domain filtering is better than spatial domain filtering, we cannot say whether it is true or false without empirical evidence. However, there are certain advantages to using frequency domain filtering for medical images, such as its ability to remove high-frequency noise without losing important details.
• To determine the impact of pre-processing steps like contrast enhancement and gamma correction on spatial and frequency domain filtering approaches, we can experiment with applying these steps before filtering and compare the PSNR values of the resulting images. It is possible that different pre-processing steps may have different effects on the two filtering techniques, so we need to analyze their performance separately.
• For the tunable parameters of pre-processing steps, we can use grid search to experiment with different values and observe how they affect the PSNR values of the filtered images. By analyzing the results, we can identify which parameter values have a positive, negative, or neutral impact on the filtering techniques.
• To determine the ideal pre-processing step and optimal value of the pre-processing tunable parameter for getting the best results of the spatial domain filtering and frequency domain filtering, we need to perform a comprehensive analysis of the various combinations of pre-processing steps and parameter values. This analysis can involve experimenting with different combinations of pre-processing steps and parameter values and comparing the PSNR values of the filtered images. By analyzing the results, we can identify the best pre-processing step and optimal parameter value for each filtering technique.

Question 2:

• In this example, img is the input image. The ORB and BRISK feature detectors are initialized using the ORB_create() and BRISK_create() functions, respectively. The detect() function is used to detect features in the image, and the compute() function is used to compute descriptors for the features. Finally, the drawKeypoints() function is used to draw the keypoints on the input image, and the resulting images with keypoints are displayed using the imshow() function.
• To transform the input image, you can use functions like cv2.warpAffine() for affine transformations (translation, rotation, scaling, shearing) and cv2.warpPerspective() for perspective transformations. Here's an example code snippet to rotate the input image by 45 degrees

• In this example, M is the transformation matrix computed using the getRotationMatrix2D() function. The warpAffine() function is used to apply the rotation to the input image, and the resulting rotated image is displayed using the imshow() function.

• To establish the performance of the feature detectors, we can use metrics like repeatability and matching score. Repeatability measures how well the same feature points can be detected in multiple images under different transformations, while matching score measures how well the detected features can be matched between multiple images. Here's an example code snippet to compute the repeatability and matching score of the ORB and BRISK feature detectors:

• To evaluate the performance of feature detectors, you can use metrics such as repeatability, matching accuracy, or descriptor distinctiveness. These metrics can be computed by comparing the features detected in different images that have undergone different transformations. For example, to evaluate the repeatability of ORB features under rotation, you can rotate the input image by different angles and compute the number of keypoints that are detected in each rotated image and in the original image. You can then plot a curve of repeatability versus rotation angle, and compare it with the curves obtained for other feature detectors or for different parameter settings of ORB.
• To optimize the performance of feature detectors, you can use preprocessing techniques such as contrast enhancement, gamma correction, or histogram equalization. These techniques can be applied to the input image before running the feature detector, and the optimal settings can be determined by grid search or other optimization methods. For example, you can apply histogram equalization to the input image with different values of the clip limit parameter, and evaluate the repeatability and distinctiveness of ORB features on the equalized images. You can then plot a curve of performance versus clip limit, and choose the optimal value that maximizes the performance.

3.

• PSNR (Peak Signal to Noise Ratio) is a measure of the quality of a processed image compared to the original image. In order to determine the combination of morphological transforms that yield the best PSNR gains, we need to first define a baseline PSNR value for the input image. We can then apply different morphological transforms to the input image and compute the PSNR values of the resulting images, and compare them to the baseline value.

• Now let's apply different morphological transforms to the input image and compare their PSNR values. We can use the following morphological transforms in OpenCV:

  • Opening: Erosion followed by dilation
  • Closing: Dilation followed by erosion
  • Top-hat: Input image minus opening
  • Black-hat: Closing minus input image

• In this example code, we apply each morphological transform to the input image using the cv2.morphologyEx() function, and compute the PSNR value of the resulting image using the psnr() function defined earlier. We then compare the PSNR values of each transformed image to the baseline PSNR value of the input image.

• Based on the results, we can determine which combination of morphological transforms yield the best PSNR gains. In this case, we combine the top-hat and black-hat transforms using the formula input_image + tophat_img - blackhat_img, which resulted in the highest PSNR gain. However, it's important to note that the optimal combination of morphological transforms may vary depending on the input image and the desired output.