Télécharger la présentation
## References

- - - - - - - - - - - - - - - - - - - - - - - - - - - E N D - - - - - - - - - - - - - - - - - - - - - - - - - - -

**References**• Book: Chapter 5, Image Processing, Analysis, and Machine Vision, Sonka et al, latest edition (you may collect a copy of the relevant chapters from my office) • Papers: • Harris and Stephens, 4th Alvey Vision Conference, 147-151, 1988. • Matas et al, Image Vision Geometry, 22:761-767, 2004**Topics**• Line detection • Interest points • Corner Detection • Moravec detector • Differential approaches • Harris corner detection • Maximally stable extremal regions**Line detection**• Useful in remote sensing, document processing etc. • Convolve with line detection kernels**?**Lines and corner for correspondence • Interest points for solving correspondence problems in time series data. • Corners are better than lines in solving the above problem due to the aperture problem • Consider that we are looking at a moving problem through a small aperture • The only component of the motion vector that is orthogonal to the line can be seen • A vertex or corner provides better correspondence**?**Corners • Challenges • Gradient computation is less reliable near a corner due to ambiguity of edge orientation • Corner detector are usually not very robust. This deficiency is overcome either by manual intervention or large redundancies. The later approach leads to many more corners than needed to estimate transforms between two two images (e.g., RANSAC for correspondence building).**Corner detection**• Moravec detector: detects corners as the pixels with locally maximal contrast • Differential approaches: • Beaudet’s approach: Corners are measured as the determinant of the Hessian. Note that the determinant of a Hesian is equivalent to the product of the min&max Gaussian curvatures**Continued …**• Kitchen and Rosenfeld approach: Corners are defined as the rate of change of gradient direction multiplied by the gradient magnitude which is equivalent to the second directional derivative in the direction orthogonal to the gradient • Using a bi-cubic facet model**Harris corner detector**• Key idea: Measure changes over a neighborhood due to a shift and then analyze its dependency on shift orientation • Orientation dependency of the response for lines Δ Δ Δ Δ Δ High response for shifts along the edge direction; low responses for shifts toward orthogonal direction direction Anisotropic response Δ**Key idea: continued …**• Orientation dependence of the shift response for corners Δ Δ Δ Δ Δ High response for shifts along all directions Isotropic response Δ**Harris corner: mathematical formulation**• An image patch or neigborhood W is shifted by a shift vector Δ=[Δx, Δy]T. • A corner does not have the aperture problem and therefore should show high shift response for all orientation of Δ. • The square intensity difference between the original and the shifted image over the neighborhood W is • Apply first-order Taylor expansion**Harris matrix**• The matrix AW is called the Harris matrix and its symmetric and positive semi-definite. PCA of AW gives an eigen vector and eigen value (λ1, λ2) system of the orientation dependency of partial derivative of local intensity function. Three distinct situations: • Both λ1 and λ2 are small no edge or corner; a flat region • λi is large but λji is small existence of an edge; no corner • Both λ1 and λ2 are large existence of a corner • Harris response function • A value of κ between 0.04 and 0.15 has be used in literature.**Algorithm: Harris corner detection**• Filter the image with a Gaussian • Estimate intensity gradient in two coordinate directions • For each pixel c and a neighborhood window W • Calculate the local structure matrix A • Compute the response function R(A) • Choose the best candidates for corners by selecting thresholds on the response function R(A) • Apply non-maximal suppression**Maximally Stable Extremal Regions (MSER)**• Key idea: Consider a movie generated by thresholding a gray level image (intensity values [0,1,…,255]) at all possible thresholds starting from 0 and ending at 255. • Initially it’s an empty image and then some dots (local minima) appear and starts growing; new dots appear and starts growing and so on; from time to time two disconnected regions get merged and finally all regions get merged into a single component. • BUT, the important observation here is that – starting from a tiny seed area (one or a few pixels), a region continues growing till it fills the object containing the initial seed area and then remains (almost) unchanged for quite sometime in the threshold movie until it get merged with the bigger (generally, parent) object to which it belongs • MSER intends to capture these stable regions**MSER: properties**• Invariance under monotonic transforms M : II of image intensities • Invariance under homeomorphic transformations (adjacency preserving) T: CC of the image space • Stability: only extremal regions remail virtually unchanged over a threshold range are selected • Multi-scale detection: Extremal regions of all scales are detected simultaneously • The set of all MSERs are enumerated in O(n*log log n) time, where n is the number of image pixels**Algorithm: MSER enumeration**• Input: Image I and the Δ parameter • Output: List of nested extremal regions • For all pixels shorted by intensity • Place a pixel in the image as its tern come • Update the connected component structure • Update the area for the effected connected components • For all connected components • Detect regions with local minima w.r.t. rate of change of connected component area with threshold; define each such region as a MSER