Multiscale Spatial Filtering Research Articles

Poster Session I: Combining spatial and quantitative models to account for the relationship between luminance, context and brightness.

The apparent brightness of a target region is determined not only by its local luminance, but also by its surrounding context. This is well demonstrated by the many illusions where some surround context causes two physically identical targets to appear different in brightness (e.g. White's (1979) illusion). Here we use perceptual scales from Maximum Likelihood Conjoint Measurements to test the predictive power of computational brightness models as a function of surround context and local luminance. While the qualitative relationship between target brightness and surround context (i.e. direction of effect) is captured well by image-computable models of the spatial filtering type (e.g. Blakeslee and McCourt, 1997), these do not predict the quantitative aspects of that relationship (i.e. magnitude of effect). A different class of models (e.g. Whittle, 1992) makes quantitative predictions of brightness as a function of luminance, but requires labeling of target and background (i.e., is not image-computable), thus cannot by itself account for the spatial context. Hence we combine Whittle's model with spatial mechanisms to account for both spatial context and local luminance. A combined model using multiscale spatial filtering and calculating brightness at every spatial scale is able to provide both a correct qualitative prediction of direction of effect, as well as a quantitative prediction of apparent brightness that can be tested against perceptual scales for different stimuli.

A history and modular future of multiscale spatial filtering models

A history and modular future of multiscale spatial filtering models

A Hybrid Multi-scale Spatial Filtering and Minimum Spanning Forest for Spectral–Spatial Hyperspectral Image Classification

Integration of spatial and spectral information is an effective way in improving classification accuracy. In this article a new framework, based on multi-scale spatial weighted mean filtering (MSWMF) and minimum spanning forest, is proposed for the spectral–spatial classification of hyperspectral images. In the proposed framework, at first the image is smoothed by MSWMF and then the first eight principal components are extracted. Using support vector machine, at each scale of MSWMF, a classification map is produced in order to generate a marker map in the next step. Then, the minimum spanning forest is built on the marker map. Finally, in order to create a final classification map, all the classification maps of each scale are merged with a majority vote rule. The experimental results of the hyper-spectral images indicate that the suggested framework enhances the classification accuracy, in comparison with previously classification techniques. So, it is interesting for hyperspectral images classification.

结合多尺度空间滤波和层级网络的基于结构保持的高光谱特征选择

结合多尺度空间滤波和层级网络的基于结构保持的高光谱特征选择

When is spatial filtering enough? Investigation of brightness and lightness perception in stimuli containing a visible illumination component

Brightness (perceived intensity) and lightness (perceived reflectance) matching were investigated in seven well-known visual stimuli that contain a visible shadow or transparent overlay. These stimuli are frequently upheld as evidence that low-level spatial filtering is inadequate to explain brightness/lightness illusions and that additional mid- or high-level mechanisms are required. The argument in favor of rejecting low-level spatial filtering explanations has been founded on the erroneous assumption that equating test patch and near surround luminance is sufficient to control for and rule out this type of mechanism. We tested this idea by comparing the matching behavior of four observers to the predictions of the ODOG multiscale filtering model (Blakeslee & McCourt, 1999). Lightness and brightness matching differed significantly only when test patches appeared in shadow or beneath a transparency. Lightness and brightness matches were both significantly larger under these conditions; however, the lightness matches greatly exceeded the brightness matches. Lightness matches were greater for test patches in shadow or beneath a transparency because lightness matches under these conditions were based on conscious inferential (not sensory-level) judgments where observers attempted to discount the difference in illumination. The ODOG model accounted for approximately 80% of the total variance in the brightness matches (as well as in the lightness matches for targets not in shadow or beneath a transparency), and successfully predicted the relative magnitude of these matches in five of the seven stimulus sets. These results indicate that multiscale spatial filtering provides a unified and parsimonious explanation for brightness perception in these stimuli and imply that higher-level mechanisms are not required to explain them. The model was not as successful for the argyle and wall of blocks illusions in that it incorrectly rank-ordered the relative magnitude of the effects across different versions of the stimuli. It is an important question whether such model failures are due to known but corrigible limitations of the ODOG model or whether they will require other (possibly higher-level) explanations.

The role of oriented filters and T-junctins in White's effect

In White's Effect the gray target surrounded more by black than white appears darker than the target of the same physical luminance surrounded more by white than black. Blakeslee and McCourt (1999) have proposed a multiscale spatial filtering (ODOG) model of White's Effect while, according to Anderson (1997), the effect is the consequence of a perceptual scission determined by the contrast polarity of aligned contours at T-junctions. As consequence of its response normalization stage, the ODOG model predicts that White's Effect should be abolished if the ratio of horizontal to vertical structure in the image is equalized. To test this prediction we constructed a variant of White's stimulus in which the background consists of zig-zag rather than straight stripes. This new stimulus is also designed to eliminate any effects of T-junctions on the difference in perceived lightness between the gray patches lying on the black and white bars. All the target patches have T-junctions at each corner, two of which are black along the top and gray/white across the stem and two of which are white across the top and gray/black across the stem. The magnitude and direction of the lightness illusion in the standard and novel White's stimuli were measured by adjusting the luminance of comparison patches on a variegated background to match the perceived lightness of the gray target patches. We found a highly significant illusion in the novel White's configuration in the same direction as the standard White's Effect. Simulations show that the ODOG filtering model predicts instead a small effect in the opposite direction. This result is also hard to reconcile with accounts based on the structure and distribution of junctions within the image.

Oriented multiscale spatial filtering and contrast normalization accounts for Howe's variation of White's effect

Oriented multiscale spatial filtering and contrast normalization accounts for Howe's variation of White's effect