Novelty detection is the identification of new or unknown data or signal that a machine learning system is not aware of during training. Novelty detection is one of the fundamental requirements of a good classification or identification system since sometimes the test data contains information about objects that were not known at the time of training the model. In this paper we provide stateof -the-art review in the area of novelty detection based on statistical approaches. The second part paper details novelty detection using neural networks. As discussed, there are a multitude of applications where novelty detection is extremely important including signal processing, computer vision, pattern recognition, data mining, and robotics.

Abstract—In this paper, we present a new framework for novelty detection. The framework evaluates neural networks as adaptive classifiers that are capable of novelty detection and retraining on the basis of newly discovered information. We apply our newly developed model to the application area of object recognition in video. This paper details the tools and methods needed for novelty detection such that data from unknown classes can be reliably rejected without any a priori knowledge of its characteristics. The rejected data is postprocessed to determine which samples can be manually labeled of a new type and used for retraining. In this paper, we compare the proposed framework with other novelty detection methods and discuss the results of adaptive retraining of neural network to recognize further unseen data containing the newly added objects. Index Terms—Scene analysis, neural networks, adaptive classifiers, novelty detection. æ

Presettlement forest types were mapped as fuzzy sets from point data representing trees contained in General Land Office survey notes (ca.1850) for Chippewa County, Michigan. The resulting representation agreed with a polygon map of the same forest types at 66 percent of the locations (represented as grid cells) in the county. Boundary vagueness was defined in relation to the slope of a linear function fitted to the negative relationship between entropy of forest types and distance to polygon boundaries. The similarity between forest type compositions (i.e., classification ambiguity) was shown to account for 55 percent of the variation in boundary vagueness.

In this paper, we introduce a set of taxonomies that hierarchically organize and specify algorithms associated with hyperspectral unmixing. Our motivation is to collectively organize and relate algorithms in order to assess the current state-of-the-art in the field and to facilitate objective comparisons between methods. The hyperspectral sensing community is populated by investigators with disparate scientific backgrounds and, speaking in their respective languages, efforts in spectral unmixing developed within disparate communities have inevitably led to duplication. We hope our analysis removes this ambiguity and redundancy by using a standard vocabulary, and that the presentation we provide clearly summarizes what has and has not been done. As we shall see, the framework for the taxonomies derives its organization from the fundamental, philosophical assumptions imposed on the problem, rather than the common calculations they perform, or the similar outputs they might yield.

Abstract not found

: A genetic approach is developed, which is suitable for the optimization of fuzzy c-means clustering. The approach is based on real encoding of the prototype variables (cluster centers) and uses appropriate genetic operators and techniques to optimize the clustering criterion. Experimental results concerning difficult clustering problems show that the proposed approach is very successful in generating fuzzy partitions and prototypes and outperforms the fuzzy c-means algorithm in terms of the correct placement of patterns into partitions. 1 INTRODUCTION The task of pattern classification and recognition typically constitutes a major component of an intelligent diagnostic system. Pattern classification can be viewed as including two steps: first, a phase of clustering given samples, and second, classification of new samples based on the knowledge of clusters. Pattern clustering considers a set of unlabeled data objects and seeks to find natural groupings amongst the exemplars. The clus...

Data on tree location and species in a portion of Northern Michigan were gathered from General Land Office (GLO) survey notes (ca. 1850), digitized, and generalized to represent forest types. Fuzzy membership values describing the degree of membership of each species in each forest type were derived from (a) semantic information in the forestry literature and (b) a fuzzy clustering routine applied to data from randomly placed circular plots. The fuzzy membership values assigned to each tree point for each forest type were interpolated to form continuous surfaces using kriging and co-kriging. Advantages of this method over traditional discrete mapping methods include: (a) multiple options are available for the display and analysis; (b) classification uncertainty and the continuity of natural vegetation can be represented; and (c) the classification scheme is applied systematically across the entire map area and can be altered to produce alternative maps. The subset of available display and analytical products presented include: discrete forest type maps; a surface representing the confusion between forest types; fuzzy logical overlays of forest types; and discrete class maps with color value altered within each class to indicate degree of confusion at each location.

A new more efficient variant of a recently developed algorithm for unsupervised fuzzy clustering is introduced. A Weighted Incremental Neural Network (WINN) is introduced and used for this purpose. The new approach is called FC-WINN (Fuzzy Clustering using WINN). The WINN algorithm produces a net of nodes connected by edges, which reflects and preserves the topology of the input data set. Additional weights, which are proportional to the local densities in input space, are associated with the resulting nodes and edges to store useful information about the topological relations in the given input data set. A fuzziness factor, proportional to the connectedness of the net, is introduced in the system. A watershed-like procedure is used to cluster the resulting net. The number of the resulting clusters is determined by this procedure. Only two parameters must be chosen by the user for the FC-WINN algorithm to determine the resolution and the connectedness of the net. Other parameters that must be specified are those which are necessary for the used incremental neural network, which is a modified version of the Growing Neural Gas algorithm (GNG). The FC-WINN algorithm is computationally efficient when compared to other approaches for clustering large high-dimensional data sets.

Abstract. The growth of cities represents huge problems for modern societies. Monitoring, analysing and modelling the urban dynamic call for detailed mapping of urban land-use. Traditionally, urban land-use mapping is based on orthophotos and satellite images, but deriving land-use from remote-sensing alone is not satisfactory. The Danish Building & Dwelling Register is a database containing detailed information like year of construction, use, area etc. Therefore, this database provides a useful foundation for urban land-use mapping. To be able to track urban land-use changes over time, we have chosen square cells (100m x 100m) as basic mapping units. Generally, land cover and land-use mapping are based on crisp classification, but in the current project we have applied a fuzzy modeling approach to land-use mapping. Fuzzy classification offers a better choice in urban land-use mapping, because it can indicate the primary, secondary etc. land-use simultaneously. This will offer more meaningful information for planners and a more detailed understanding of the land-use patterns. Based on these principles, a nation wide urban land-use database for the year 1997 is established. 1

Abstract. A simple, efficient algorithm is presented for super-resolution target mapping from remotely sensed images. Following an initial random allocation of pixel proportions to binary 'hard ' sub-pixel classes, the algorithm works in a series of iterations, each of which contains two stages. For each iteration, a distance weighted function of neighbouring pixels is computed for all sub-pixels. Then, on a pixel-by-pixel basis, the '1 ' with the minimum value of the function is swapped with the '0 ' with the maximum value of the function, if the swap results in an increase in some objective function. The algorithm is demonstrated to work reasonably well with simple images, opening the way for further research to explore the algorithm, to extend the algorithm to multiple classes and to develop more efficient, but equally simple algorithms. 1.