[Group Home]   [Research Statement]   [Projects]   [Publications]   [Members]  

Research Projects

[Texture Modeling]   [Texture Classification]   [Face Detection]  [Optimal Componet Analysis]  [Joint Shape-Texture Models] 

D084_128_julesz_sparse_zhu_small_seq Texture Modeling and Synthesis

New! Results_FSvision_new_crop Face detection using spectral histograms

New! sealion_morphed_dolphinJoint shape-texture models

New! xor_basis_517_images_aniOptimal Component Analysis

Research projects at the Florida State Vision group centers around the visual image modeling and visual inference with emphasis on its applications. Several of our current research projects utilize a visual feature called Spectral Histograms. A spectral histogram consists of marginal responses of a chosen set of filters, or histograms. As colors are responses at different wavelengths, visual patterns are responses of different filters. Analogy to color spectrum, we name the visual feature Spectral Histograms. Our extensive study has shown the spectral histogram exhibits very nice properties that are consistent with human visual perception. For example, it exhibits a nonlinearity which matches the psychophysical data on texture perception very well. The following table shows four texture images along with its spectral histogram of 8 filters. Each texture has a unique spectral histogram. Similar textures tend to have similar spectral histograms as shown in these examples.
Texture ImageSpectral Histogram
/step_different_m.pgm_3000_1000_obs step_spectral_hist
crosses.pgm_obs cross_spectral_hist
Fabric_0009_texture_julesz_obs Fabric_0009_spectral_hist
D084_sparse_zhu_filter_obs_0 D084_spectral_hist
We can show that the spectral histogram of a texture captures the essential regularities and visual structures by synthesizing similar texture patterns through matching spectral histograms using Monte Carlo Markov Chain methods. Note that circular boundaries are used in all the examples. The first pair is quite similar under the circular boundary condition.
Texture ImageSynthesized Textures
step_different_m.pgm_3000_1000_obs step_different.pgm_syn_ver_3500
crosses.pgm_obs crosses.pgm_syn_ver_4000
Fabric_0009_texture_julesz_obs Fabric_0009_texture_julesz_julesz_syn_final_55
D084_sparse_zhu_filter_obs_0 D084_sparse_zhu_filter_syn_147
The spectral histogram essentially gives a way to overcome many problems in texture modeling, classification, and segmentation. This also gives a way to model human visual perception which can match psychophysical data as well as quantatively explain some of the perceptual phenomena. Please visit each project's description for more details.
[Group Home]   [Research Statement]   [Projects]   [Publications]   [Members]  

Last modified on June 10, 2002.