Context-dependent fusion with application to landmine detection.

Abstract

Traditional machine learning and pattern recognition systems use a feature descriptor to describe the sensor data and a particular classifier (also called "expert" or "learner") to determine the true class of a given pattern. However, for complex detection and classification problems, involving data with large intra-class variations and noisy inputs, no single source of information can provide a satisfactory solution. As a result, combination of multiple classifiers is playing an increasing role in solving these complex pattern recognition problems, and has proven to be viable alternative to using a single classifier. In this thesis we introduce a new Context-Dependent Fusion (CDF) approach, We use this method to fuse multiple algorithms which use different types of features and different classification methods on multiple sensor data. The proposed approach is motivated by the observation that there is no single algorithm that can consistently outperform all other algorithms. In fact, the relative performance of different algorithms can vary significantly depending on several factions such as extracted features, and characteristics of the target class. The CDF method is a local approach that adapts the fusion method to different regions of the feature space. The goal is to take advantages of the strengths of few algorithms in different regions of the feature space without being affected by the weaknesses of the other algorithms and also avoiding the loss of potentially valuable information provided by few weak classifiers by considering their output as well. The proposed fusion has three main interacting components. The first component, called Context Extraction, partitions the composite feature space into groups of similar signatures, or contexts. Then, the second component assigns an aggregation weight to each detector's decision in each context based on its relative performance within the context. The third component combines the multiple decisions, using the learned weights, to make a final decision. For Context Extraction component, a novel algorithm that performs clustering and feature discrimination is used to cluster the composite feature space and identify the relevant features for each cluster. For the fusion component, six different methods were proposed and investigated. The proposed approached were applied to the problem of landmine detection. Detection and removal of landmines is a serious problem affecting civilians and soldiers worldwide. Several detection algorithms on landmine have been proposed. Extensive testing of these methods has shown that the relative performance of different detectors can vary significantly depending on the mine type, geographical site, soil and weather conditions, and burial depth, etc. Therefore, multi-algorithm, and multi-sensor fusion is a critical component in land mine detection. Results on large and diverse real data collections show that the proposed method can identify meaningful and coherent clusters and that different expert algorithms can be identified for the different contexts. Our experiments have also indicated that the context-dependent fusion outperforms all individual detectors and several global fusion methods.