In this paper, a new scene recognition visual descriptor called Enhanced Scale Invariant Feature Transform-based Sparse coding Spatial Pyramid Matching (Enhanced SIFT-ScSPM) descriptor is proposed by combining a Bag of Words (BOW)-based visual descriptor (SIFT-ScSPM) and Gist-based descriptors (Enhanced Gist-Enhanced multichannel Gist (Enhanced mGist)). Indoor scene classification is carried out by multi-class linear and non-linear Support Vector Machine (SVM) classifiers. Feature extraction methodology and critical review of several visual descriptors used for indoor scene recognition in terms of experimental perspectives have been discussed in this paper. An empirical study is conducted on the Massachusetts Institute of Technology (MIT) 67 indoor scene classification data set and assessed the classification accuracy of state-of-the-art visual descriptors and the proposed Enhanced mGist, Speeded Up Robust Features-Spatial Pyramid Matching (SURF-SPM) and Enhanced SIFT-ScSPM visual descriptors. Experimental results show that the proposed Enhanced SIFT-ScSPM visual descriptor performs better with higher classification rate, precision, recall and area under the Receiver Operating Characteristic (ROC) curve values with respect to the state-of-the-art and the proposed Enhanced mGist and SURF-SPM visual descriptors.

Support vector machine (SVM) methods are widely used for classification and regression analysis. In many engineering applications, only one class of data is available, and then one-class SVM methods are employed. In reliability applications, the one-class data may be failure data since the data are recorded during reliability experiments when only failures occur. Different from the problems handled by existing one-class SVM methods, there is a bias constraint in the SVM model in this work and the constraint comes from the probability of failure estimated from the failure data. In this study, a new one-class SVM regression method is proposed to accommodate the bias constraint. The one class of failure data is maximally separated from a hypersphere whose radius is determined by the known probability of failure. The proposed SVM method generates regression models that directly link the states of failure modes with design variables, and this makes it possible to obtain the joint probability density of all the component states of an engineering system, resulting in a more accurate prediction of system reliability during the design stage. Three examples are given to demonstrate the effectiveness of the new one-class SVM method.

We propose a feature selection method for density estimation withquadratic loss. This method relies on the study of unidimensionalapproximation models and on the definition of confidence regions forthe density thanks to these models. It is quite general and includescases of interest like detection of relevant wavelets coefficientsor selection of support vectors in SVM. In the general case, weprove that every selected feature actually improves the performanceof the estimator. In the case where features are defined bywavelets, we prove that this method is adaptative near minimax (upto a log term) in some Besov spaces. We end the paper bysimulations indicating that it must be possible to extend theadaptation result to other features.


Structural optimization is usually handled by iterative methods requiring repeated samples of a physics-based model, but this process can be computationally demanding. Given a set of previously optimized structures of the same topology, this paper uses inductive learning to replace this optimization process entirely by deriving a function that directly maps any given load to an optimal geometry. A support vector machine is trained to determine the optimal geometry of individual modules of a space frame structure given a specified load condition. Structures produced by learning are compared against those found by a standard gradient descent optimization, both as individual modules and then as a composite structure. The primary motivation for this is speed, and results show the process is highly efficient for cases in which similar optimizations must be performed repeatedly. The function learned by the algorithm can approximate the result of optimization very closely after sufficient training, and has also been found effective at generalizing the underlying optima to produce structures that perform better than those found by standard iterative methods.