Scribd
Upload
12 views4 pages

Lab3 110061203

Uploaded by

yuqiwang677
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
12 views4 pages

Lab3 110061203

Uploaded by

yuqiwang677
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
You are on page 1/ 4

DSP 實驗 110061203 王譽錡 Wang,Yu-Chi

Lab3

1.

In the first image (data/Im.jpg), the corner detection results are expected to be sharp
due to the presence of strong edges and clear textures. Gaussian smoothing
effectively reduces noise, allowing corners to be detected at junctions where
significant intensity changes occur. For the second image, which may contain more
texture or complex patterns, the corner detection could reveal a greater number of
corner points. The Gaussian filter plays a key role in minimizing noise interference,
ensuring that corners are not falsely detected in these textured areas. In contrast, the
third image, characterized by smoother regions, will likely yield fewer detected
corners. The Gaussian smoothing, in this case, suppresses false corner detections in
uniform areas, leading to a cleaner output.
Comparing the results across images, Gaussian smoothing is observed to smooth out
the gradient product IxIyI_x I_yIxIy, making corner detection less sensitive to noise
and more focused on regions with significant corner-like structures. The final corner
detection results show strong corner features as indicated by the Harris response
function. As image complexity increases, with more textures or sharp features, a
higher number of corners are detected. This visualization of multiple images
demonstrates how varying image textures and content impact the performance and
accuracy of corner detection.
3.

The box filter tends to blur the image, and it doesn't handle edges well because it
introduces pixels that look like box-like shapes, and the smoothing is not smooth in a
natural way. The uniform weight applied across the kernel can lead to image in block,
and it make the transition between high and low frequencies quite not apparent.
The Gaussian filter’s smoothing effect is more visually-pleasing, making the
transitions between high and low frequencies smoother and less noticeable. This
leads to visually-pleasing hybrid images. The edge preservation of Gaussian filters
tend to preserve boundaries better than box filters because of their more fitting
weighting coefficient.
All in all, I think the basis of the reason is that the box filter is manipulating by
simple averaging. It gives equal weight to all pixels in the neighborhood, regardless
of their distance from the center, so it will not handle the edges as well as the
Gaussian. And the box filter produce the image like the Gaussian filter with a middle
sigma. The high frequency part of the cat is clear to see.
4.
For the first flaw is the increasing sensitivity to noise. Without smoothing, the
gradients Ix and Iy will contain high-frequency noise present in the image. This noise
gets amplified in the second moment matrix components, leading to noisy corner
responses.
As a result, many false corners (spurious detections) might be detected, especially in
areas of the image with small texture variations or pixel-level noise.
Failure to Detect Corners in Complex Areas, in textured or complex regions, the
second moment matrix without smoothing might not capture the overall gradient
structure well. This can lead to poor corner detection or failure to identify corners in
regions where the gradient changes are subtle but significant.
And I put the step away, it create some error corner points, and for the reason of
spatial discontinuities, the Smoothing helps ensure that the detected corner points are
spatially consistent and well-localized. Without it, the algorithm might produce
corner detections that are less spatially coherent, leading to erratic or clustered corner
points.

5.

Differences and Similarities Between Prewitt and Scharr Gradient Filters


Their differences lies in accuracy, noise sensitivity, gradient estimation. From the
perspectives of accuracy, the Scharr filter is more accurate in detecting
gradients and as a result, produces more precise corners and edges. Prewitt is
less sensitive and tends to miss subtle intensity changes. Noise Sensitivity: Scharr
is more robust against noise due to its higher weights in the filter, while Prewitt
can be more affected by noise, leading to less reliable corner detection.Gradient
Estimation: Scharr provides better gradient approximation, especially for
rotations and high-frequency details, making it superior for detecting fine image
details.
But they are still some similarities
 Purpose: Both are first-order derivative filters that estimate the gradient of the
image's intensity to detect corners.
 Size and Efficiency: Both use 3x3 kernels, making them computationally
efficient. However, Scharr achieves better accuracy at the same kernel size.
 Usage in Edge and Corner Detection: Both filters are commonly used in
edge and corner detection, but Scharr is typically favored for tasks requiring
higher precision.

result image:

Prewitt filter :

Scharr filter :

You might also like