How Can I Improve Mask Reconstruction Accuracy in Image Sequences?

[wizards]

Dear Sir,

I am doing a project in VC++. I have a sequence of images which are scanned from video tape.

I need to apply some filters locally to these images. For this I will start with the first image of the sequence.

I will mask an area (like face, hand of human) which I need to apply the filter. Now I will apply the filter on this mask (I stored all the pixels of the mask in a structure).

Now my requirement is to apply the same effect for the same resign (mask) in the remaining images. That is without masking the resign again I need to apply the same filter in the next images.

I tried this with SIFT. I created SIFT features for mask and for the entire targeted image (next image in the sequence). I calculated Euclidean distance of these two. I got some I got some similar points.

Now I segmented the target image and collected all pixels of the segment, if there is a matching point in this segment. This I am doing for all the matching points.

The problem with this approach is it is taking more time to calculate SIFT features and segmenting. For a mask of size 100x100 over an image is taking more than 40 seconds.

And if there is any rotation/scaling in the mask in the next image then it is not reconstruction the mask perfectly. In my case it gives 30% accuracy. That is some borders are bleeding.

Can you please help me what to do to get more accuracy with less time.

Can you please suggest me the correct algorithm if SIFT is not perfect.

I can send you my work space if required.

Thanks in Advance

Hema Lak

 

[lofgran]

shmi

Dear Hema Lakshmi,

Thank you for sharing your project with us. It sounds like you are trying to apply a filter to a specific area in a sequence of images, and you are currently using the SIFT algorithm for this task. While SIFT is a popular and effective algorithm for feature matching, it may not be the best choice for your specific project.

One alternative approach you could consider is using a convolutional neural network (CNN) for feature extraction and filtering. CNNs are commonly used in image processing tasks and can provide accurate results with less computational time compared to SIFT. You can train the CNN on a set of images with the desired filter applied to the specific area, and then use it to apply the same filter to the rest of the images in the sequence.

Another suggestion is to use a different feature matching algorithm, such as SURF or ORB, which may provide better results for your specific task. These algorithms also have the advantage of being faster than SIFT.

Lastly, you could also try implementing a template matching algorithm, which involves comparing a template image (your mask) to the rest of the images in the sequence to find the best match. This approach may be more accurate and faster than using feature matching algorithms.

I hope these suggestions are helpful to you. If you would like further assistance, please feel free to share your workspace with us and we will be happy to take a closer look at your project.