01
Image Processing
Raw pixels are filtered, enhanced, denoised, and prepared for higher-level analysis.
Ofer Miller, Ph.D.
Computer Vision . Image processing . Graph Theory Models . Video Analysis
Ph.D. in Computer science | Researcher in Computational Imaging & Signal analysis
Main Menu
*Internal Links
*External Links
HONORS AND AWARDS
• "Celia and Marcos Maus" Annual Prize” in Computer Science for outstanding achievements in doctoral research, May 2002.• Excellence Scholarship for Doctoral Students from the Council for Higher Education in the field of high technology, December 2000• Tel Aviv University Award for Excellence in Master’s Studies in Science, June 1999.• Tel Aviv University Scholarship for outstanding Master’s Students in Science, October 1998
Publications (Contribute to Science)
List of Patents
| Title | Author | Year | Patent#: |
|---|---|---|---|
| 1. "Determine Viewer's exposer to Visual Messages" | Ofer Miller | 2016 | 20170169464 |
| 2. "Method for logging a user in to a mobile device" | Ofer Miller | 2015 | 20150049922 |
| 3. Method for rating areas in video frames | Ofer Miller... | 2013 | 8457402 |
| 4. Method and Device for Processing Video Frames | Ofer Miller... | 2012 | 20120017238 |
| 5. System and Method for Enriching Video Data | Ofer Miller... | 2011 | 20110217022 |
| 6. Method for illumination independent change detection in a pair of registered gray images | Ofer Miller... | 2006 | 7088863 |
| 7. Automatic object extraction | Ofer Miller... | 2006 | 7085401 |
Research contribution during studies at TLV University
Illumination independent Object based Change Detection Article
Video Graph based Segmentation of Moving Objects while spatial and temporal information are available.
Still Image Segmentation Based on Adaptive Local Thresholds
Tracking of Moving Objects Based on Matching between Graph Edges
Contribution to Artimedia Initiative
Foundations of Computer Vision
Computational Imaging & Graph-Based Vision — Dr. Ofer Miller
Miller's Research Publications
Segmentation
Change Detection
Graph Theory
Face Recognition
Tracking
Obj Segmentation
Miller Perspective on AI
From K-Means to AI
A useful starting point for understanding artificial intelligence is the K-means algorithm, a classical unsupervised machine-learning method that groups data points into k clusters according to their similarity. Each point is assigned to the cluster whose centroid is nearest, and the centroids are then recomputed as the mean of their assigned points. This assignment-and-update process is repeated until the cluster memberships stabilize.
Although K-means is relatively simple, it captures a fundamental computational principle: complex data can be organized into structured groups by iteratively improving an internal representation. This makes it useful in tasks such as customer segmentation, image analysis, and visual-data organization. Its limitations are also instructive: the number of clusters must be specified in advance, and the method is most effective when the underlying clusters are approximately spherical in Euclidean space.
Modern artificial intelligence systems operate by learning statistical patterns from data and using those patterns to make predictions, classifications, decisions, or generative outputs. Deep neural networks extend this idea by optimizing large collections of parameters through iterative training. The model compares its predicted output with the desired output, computes an error, and adjusts its internal weights using gradient-based optimization methods such as backpropagation.
Conceptually, deep learning may be viewed as a vast generalization of the intuition embodied in K-means. Whereas K-means partitions data in a relatively simple geometric space, deep neural networks learn high-dimensional feature spaces in which similar inputs are mapped to nearby regions and dissimilar inputs are separated. In this sense, AI systems perform a dynamic, learned, and highly expressive form of pattern organization, preserving the core idea of grouping similarity while expanding it toward abstraction, generalization, reasoning, and complex decision-making.
Conceptual Evolution
From Clustering to Learned Representation
01
K-Means Clustering
Data points are grouped around nearest centroids in a geometric feature space.
02
Feature Organization
Patterns become compact internal structures that reduce raw complexity.
03
Deep AI Models
Neural networks learn high-dimensional representations through optimization.
Basic and Visual Terms
Segmentation
Is it possible to segment "correctly" on still Images ? (while no motion info available)
Here is the official publication Link of the above article.
The published algorithm was considered to be the Most Cited paper !
Moving Objects Segmentation
Graph based Segmentation of Moving Objects Based on Connectivity Analysis of Spatio-temporal information
Here is the official publication Link of the above article.
Change Detection
Identify changes between two gray images taken in extreme! different illumination.
Here is the publication Link of the above article.
Smart Tracking
AI & Computer Vision
What is the AI algorithm ? how its evolved ?
Lets start with understanding of what is The K-means algorithm: so its an unsupervised machine learning algorithm used to group data points into k clusters based on their similarity, with each data point assigned to the cluster with the nearest centroid (cluster center). It works by iteratively assigning data points to clusters and then updating the centroids to the mean of the assigned points, repeating the process until no data points change cluster membership. K-means is useful for applications like customer segmentation and image analysis but requires the number of clusters (k) to be specified beforehand and is best at finding spherical clusters.
Artificial intelligence systems operate by learning statistical patterns from data and using those patterns to make predictions or decisions. At their core, modern AI models, especially deep neural networks, optimize a large set of parameters through iterative training, where the algorithm compares its predicted output to the correct output, computes an error, and adjusts internal weights using gradient-based optimization methods such as backpropagation. Through repeated exposure to vast amounts of labeled or unlabeled data, the model gradually minimizes this error and forms a high-dimensional representation of the underlying structure of the problem. Once trained, the AI system applies these learned representations to new inputs, enabling tasks such as classification, pattern recognition, reasoning, and generative output.
Now , lets jump all the way trying to describe the link between the basic k-means concept and the AI algorithms :
So the k-means algorithm, while not a learning model in the sense of deep neural networks, is a foundational unsupervised machine-learning method that contributes to the broader architecture of artificial intelligence. Through an iterative process of assigning data points to the nearest centroid and recomputing those centroids, k-means minimizes intra-cluster variance and produces a compact representation of complex datasets. In the context of AI, k-means plays the main role as a preprocessing or feature-extraction, and representation-learning tool. It is used to reduce dimensionality before training neural networks, to initialize parameters in deep models. More broadly, k-means embodies the principle that AI systems often learn by organizing and compressing information in structured ways, enabling models to generalize from raw data to meaningful abstractions.
so , what we have here ??, conceptually, AI particularly is a deep learning, can be viewed as a vast generalization of the principles embodied in k-means clustering. while k-means partitions data in a relatively simple Euclidean space by assigning each point to the nearest centroid, deep neural networks expand this notion into extremely high-dimensional, learned feature spaces. In these spaces, the model implicitly performs a form of dynamic clustering: representations of similar inputs are projected into nearby regions, while dissimilar inputs are pushed apart. Thus, AI systems can be interpreted as applying k-means-like separation but across hundreds or thousands of learned dimensions, with the boundaries not fixed by geometry alone but continually adjusted through gradient-based optimization. In this sense, the core intuition of k-means grouping similar patterns is preserved, while the expressive capacity is vastly expanded to support abstraction, generalization, and complex decision-making. so , just for those who like visual , here its how its look like :
Graph Model Representation
Ultimate Face Recognition
