prof. Ing. Michal Haindl, DrSc.

Publikace

Multispectral Texture Benchmark

Autoři
Kříž, P.; Haindl, M.
Rok
2023
Publikováno
Procedia Computer Science. 2023, 225 3143-3152. ISSN 1877-0509.
Typ
Článek
Anotace
Dozens of textural features have been published, but their realistic validation for efficient recognition applications still needs to be discovered. Textural features are derived using various approaches. We present a benchmark that can be used to evaluate these features and categorize them based on their information efficiency. We propose how the features can be benchmarked and explain different ways of measuring their properties and performance. Most textural feature-extracting algorithms are only based on information extraction from monospectral images (gray-level). Apart from native multispectral algorithms, we generalize some of these originally monospectral features for hyperspectral textures in our illustrating examples.

Texture Quality Criteria Comparison

Autoři
Haindl, M.; Shaikh, N.
Rok
2023
Publikováno
2023 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops. New Jersey: IEEE, 2023. p. 1-5. ISBN 979-8-3503-0261-5.
Typ
Stať ve sborníku
Anotace
Visual scene recognition or modeling predominantly uses visual textures representing an object's material properties. However, the single material texture varies in scale and illumination angles due to mapping an object's shape. We present a comparative study of thirteen possible texture quality criteria and show the superior performance of two multispectral measures derived from the Markovian descriptive model.

Dynamic Texture Similarity Criterion

Rok
2018
Publikováno
2018 24rd International Conference on Pattern Recognition (ICPR). Piscataway, NJ: IEEE, 2018. p. 904-909. ISSN 1051-4651. ISBN 978-1-5386-3788-3.
Typ
Stať ve sborníku
Anotace
Evaluating the likeness of two similar dynamic textures or dynamic textures of the same type is still a challenging and unresolved problem. Temporal dimension and dynamics in texture complicates the problem of mere texture similarities and makes it even more challenging. A simple approach to compute difference between DTs bag of features, STSIM of pure statistisc methods are insufficient and does not affect the variable texture dynamics which is crucial for human recognition. Testing the similarity, quality and fidelity of both natural and arftificial dynamic textures is a problem that can be solved by psychometric tests with users, but these are challenging both in terms of time, human resources and data processing. The solution provided by us compares the frequency and regularity of the time behavior of spatial frequencies in texture with great consistency with the values provided by users testing. The solution itself provides a functional metric that can be used to evaluate the similarity of textures modified by inpainting, retouching as well as evaluating the similarity of the dynamics across the type of the DTs.

Dynamic Texture Editing

Rok
2015
Publikováno
Proceedings of the 31st Spring Conference on Computer Graphics. New York: ACM, 2015. p. 133-140. ISBN 978-1-4503-3693-2.
Typ
Stať ve sborníku
Anotace
A fast simple method for dynamic textures enlargement and editing is presented. The resulting edited dynamic texture is a mixture of several color dynamic textures that realistically matches the given color textures appearance and respects their original optical flows. The method simultaneously allows to spatially and temporarily enlarge the original dynamic textures to fill any required four dimensional dynamic texture space. The method is based on a generalization of the prominent static double toroid-shaped texture modeling roller method to the dynamic texture domain. The presented method keeps the original static texture roller principle of separated analysis and synthesis parts of the algorithm. In its analytical step, the input textures patches are found by an optimal overlap tiling and the subsequent minimum boundary cut. The optimal toroid-shaped dynamic texture patches are created in each spatial and time dimension, respectively. The spatial dimension tile border is derived by textural features, color-tone, and the minimal overlapping error. The time dimension tile border is detected by minimizing the overlapping error and using the input textures optical flow. The realistic appearance of the dynamic textures mix requires to edit the patch color space and to find border patches which consists from more than one type of the texture. These border patches are found similarly to the multi-texture analysis patch step. Since all time-consuming processing, such as the finding of optimal spatio-temporal triple toroidal patches, are done in the analytical step which is completely separated from synthesis part, the synthesis of the edited and enlarged resulting texture can by done very efficiently by applying simple set of repeating rules for these triple toroidal patches. Thus the presented method is extremely fast and capable to synthesize a learned natural dynamic texture spatially and its time span in real-time.

Dynamic Texture Enlargement

Rok
2013
Publikováno
29th Proceedings Spring Conference on Computer Graphics 2013. Bratislava: Comenius University, 2013, pp. 13-20. ISBN 978-80-223-3377-1. Available from: http://www.sccg.sk
Typ
Stať ve sborníku
Anotace
A simple fast approach for dynamic texture synthesize that realistically matches given color or multispectral texture appearance and respects its original optic flow is presented. The method generalizes the prominent static double toroid-shaped texture modeling method to the dynamic texture synthesis domain. The analytical part of the method is based on optimal overlapping tiling and subsequent minimum boundary cut. The optimal toroid-shaped dynamic texture patches are created in each spatial and time dimension, respectively. The time dimension tile border is derived from the optical flow of the modeled texture. The toroid-shaped tiles are created in the analytical step which is completely separated from the synthesis part. Thus the presented method is extremely fast and capable to enlarge a learned natural dynamic texture spatially and temporally in real-time.

Efficient Textural Model-Based Mammogram Enhancement

Autoři
Haindl, M.; Remeš, V.
Rok
2013
Publikováno
IEEE 26th International Symposium on Computer-Based Medical Systems (CBMS). Los Alamitos: IEEE Computer Society, 2013, pp. 522-523. ISBN 978-1-4799-1053-3. Available from: http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=6627859&tag=1
Typ
Stať ve sborníku
Anotace
An efficient method for X-ray digital mammogram multi-view enhancement based on the underlying two-dimensional adaptive causal autoregressive texture model is presented. The~method locally predicts breast tissue texture from multi-view mammograms and enhances breast tissue abnormalities, such as the sign of a developing cancer, using the estimated model prediction error. The~mammogram enhancement is based on the cross-prediction error of mutually registered left and right breasts mammograms or on the single-view model prediction error if both breasts' mammograms are not available.

Non-Iris Occlusions Detection

Autoři
Haindl, M.; Krupička, M.
Rok
2013
Publikováno
2013 IEEE Sixth International Conference on Biometrics: Theory, Applications and Systems (BTAS). Piscataway: IEEE, 2013, pp. 1-6. ISBN 978-1-4799-0527-0. Available from: http://www.btas2013.org/
Typ
Stať ve sborníku
Anotace
The prerequisite for the accurate iris recognition is to detect all iris occlusions which would otherwise confuse a recognition method and impair its recognition rate. This paper presents a fast multispectral eyelid, eyelash, and reflection detection method based on the underlying three-dimensional spatial probabilistic textural model. The model first adaptively learns its parameters on the flawless iris texture part and subsequently checks for non iris occlusions using the recursive prediction analysis. We provide colour iris occlusion detection results that indicate the advantages of the proposed method and compare it with 97 recent Noisy Iris Challenge Evaluation algorithms.

Potts Compound Markovian Texture Model

Autoři
Haindl, M.; Remeš, V.; Havlíček, V.
Rok
2012
Publikováno
Proceedings of the 21st International Conference on Pattern Recognition. Piscataway, NJ: IEEE, 2012. p. 29-32. ISSN 1051-4651. ISBN 978-4-9906441-0-9.
Typ
Stať ve sborníku
Anotace
This paper describes a novel multispectral parametric compound Markov random field model for texture synthesis. The proposed compound Markov random field model connects a parametric control random field represented by a hierarchical Potts Markov random field model with analytically solvable wide-sense Markovian representation for single regions. The compound random field synthesis combines the modified fast Swendsen-Wang Markov Chain Monte Carlo sampling of the hierarchical Potts MRF part with the fast and analytical synthesis of single regional MRFs.