This paper introduced a method to solve low-level vision problems, such as image restoration and segmentation, optical flow, edge detection…etc. Apply the Markov random field to model the generative probability with statistical approach. Then it automatically learns relationships between observation image and corresponding scene. Finally we can use the posterior to get the implied parameters.
The goal of this work is face detection using boosting algorithm. There are major contributions of this paper. First, a new image representation called an integral image, which is similar as Haar filters. It allows us to do very fast feature evaluation. The second is taking the whole haar-like filter feature as the candidates for feature selection. And apply them to train a classifier by selecting a small number of important features using AdaBoost. The third major contribution is a method for combining successively more complex classifiers in a “cascade” why to increases the speed of the detector. It focuses attention on promising regions of the images. The final face detector is constructed with a sequence of classifiers, each slightly more complex and if any classifier rejects the sub-window, no further processing is performed.
What is Spectral Algorithm? Broadly speaking, any algorithm used in the SVD (singular value decomposition) called Spectral Algorithm. From the PCA / LDA, to the more recent Spectral Embedding / Clustering, all can be seen as a spectral algorithm. So why use SVD? The reason is some research problems are the NP-hard. To find a better approximation algorithm will spend a lot of effort. Even find a polynomial approximation, will also have issues such as too slow or encounter local minima. In contrast, SVD theory has one solution only and the algorithm is relatively faster. Moreover, SVD has many useful properties.
With graph model expansion to segmentation problem, what is good image segmentation? Generally speaking, we can see the all pixels in an image as a vertex set. The edges between nodes are assigned to their corresponding similarity (such as color, brightness, texture…). And we want to cluster the vertex set into several disjoint subset. Even more, we expect that the sum of weighted edge between the disjoint sets is small, which imply the high intra-difference. And also expect on low inter-difference. They are two issues considered basically.
Finding the “min-cut” is a well-studied problem for partition. However, the minimum cut criteria favors cutting small sets of isolated nodes in the graph. To avoid this unnatural bias, they proposed normalized cut, which simply divide the weighted sum by group association to all nodes in a normalized way.
This paper introduces a method for face images annotation with inaccurately and ambiguously labels. The different from dace recognition research before is to distinguish between faces. Its key is to identify discriminate coordinates. For this propose, it begin with doing discriminant analysis first by two methods, kernel principal components analysis (kPCA) and linear discriminant analysis (LDA). The kPCA is used to reduce the dimensionality of the data. And then LDA process is suited for the discrimination task. The two method above both used common in dimensional reduction problem.
Indexing is getting more and more important nowadays due to the increasing size of data, information or documents. Tragically, the difficulty of solving nearest neighbor-searching efficiently grows rapidly with dimension. But hopefully, researches have shown that if we can loosen the requirement of finding the exact nearest neighbor, we can still get the qualified results with negligible degradation.
In general, the purpose of using LDA is the same as PLSA – find latent topics of datasets. It can be seem as the enforced version of PLSA. LDA claims there are some problems of PLSA. First, linear increase with M. Second, Over-fitting problem. Third, PLSA can’t decide new document not in the training set (not generative).
At the first, PLSA is used to solve vocabulary mismatch problem. The work introduces a concept "latent semantic" which means the vague meaning of a cluster document which may contain co-occurrence words. It’s the LSA spirit. Compare to original LSA, PLSA is based on a mixture decomposition derived from a latent class model. This results in a more principled approach which has a solid foundation in statistics.
The definitions of "Textons" remain a vague concept still. In the early Psychophysics views, Textons refer to fundamental micro-structure in natural images (and videos) are considered as the atoms of pre-attentive human visual perception. About the Segmentation and recognition problems, Textons are very useful for such problems.
Though a good shape presentation of image is useful for image retrieval problem, there are still uncontrollable property of it. It remained difficult research problem these years. A good shape presentation should have property that invariance of scale, transition and rotation. Seldom methods can achieve all of them and present a shape clear and well. In our lecture in ammai, we survey two of the representative works of “shape”.
