I am dealing with massive 3D binary images data sets of approximately 1000**3 to 2000**3 voxels. Currently, I am using scikit-image watershed segmentation implementation to get segmented regions from input images. The algorithm consumes high amount of computational memory approximately 7 times more than what is required by the input image, so it becomes infeasible to segment big datasets. Is there any segmentation technique that consumes less RAM or any other implementation of watershed segmentation in python?
I have already tried ImageJ, Mahotas, SMIL, ITK, OpenCV and none of them worked for me. Any help in this regard will be appreciable. Thanks
I recommend considering domain decomposition via a sliding window. Iteratively take a sub-sample of the large image, segment it, and save it in the output matrix.
To make it more precise, you can select the sub-domains to be overlapped and then average out the common sections.
U can use soft computing algorithms as they are computationsly slow but are accurate
I suggest you may first consider PCA to reduce the overall size of your data. It will enable your algorithm to run much faster. Also you may use Bayesian method for segmentation, e.g., https://arxiv.org/pdf/1502.01400.pdf and http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.65.4400&rep=rep1&type=pdf
https://scikit-learn.org/stable/modules/computing.html#parallelism-resource-management-and-configuration
Thanks to all for their suggestions.
I recently wrote an article that parallelizes the watershed segmentation of porous material via geometric domain decomposition. This is effective for the domain sizes I am currently using. The algorithm is opensource and written in Python, so anybody can utilize it for his purpose. I am sharing the paper link here so that anyone may get benefit in the future.
Article Efficient extraction of pore networks from massive tomograms...
Article Computation and Memory Efficient Image Segmentation
https://scikit-learn.org/stable/modules/computing.html#computational-performance
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Computer Graphics and Animation
- 3D