Feature Extraction Using Information-Theoretic Learning
IEEE Transactions on Pattern Analysis and Machine Intelligence
| Hi-index | 0.00 |
Multistage automatic target recognition (ATR) algorithms provide an effective means for locating items of interest in sonar imagery by successively reducing the volume of data at each stage in order to produce a concise list of possible target locations. Traditionally, the stages in these algorithms are designed using statistical methods that are often based on a priori assumptions about the input data distributions. By extending recent advances in information-theoretic learning (ITL), this dissertation applies concepts of information theory to the feature extraction stage of the multistage sonar ATR paradigm. By incorporating information-theoretic performance metrics, the approach is not dependent on a priori statistical assumptions about the data distributions. The widespread application of ITL has likely been hindered by two phenomena: computational complexity and difficulty of use. This dissertation addresses both of these issues in order to increase the applicability of ITL. First, the computation of the ITL criteria is optimized for more efficient evaluation. Next, training efficiency is improved by tailoring some of the most popular advanced training algorithms for compatibility with ITL systems. To further improve training efficiency, a batch training approach to ITL is presented that reduces the quadratic complexity of the criterion. Finally, various properties of the information-theoretic criteria are investigated in order to achieve a better understanding of the ITL process in support of easier implementation. These advances in ITL efficiency and understanding are then applied to address the problem of sonar feature extraction for ATR. This problem represents a difficult real-world challenge for the application of ITL and is representative of the class of problems that stand to benefit from a more widespread implementation of ITL principles. The information-theoretic feature extraction methods presented are shown to be more effective and robust than the most popular statistical methods and some of the currently fielded sonar ATR algorithms. The ITL methods presented provide an alternative approach for designing the subspace mapping functions that are prevalent throughout ATR systems.