The application of evolutionary algorithms in the search of relevant statistical features of infrared images

A thermographic image can be a source of diagnostic information. This information can be obtained using a variety of image analysis methods. Unfortunately, informational noise resulting from the large number of features can cause problems with the efficient assessment of object state. There are, however, methods which allow one to search for relevant features useful in diagnostics. In this paper an application of evolutionary algorithms (EAs) is presented for selection of optimal (from diagnostics point of view) a set of statistical features of infrared images. The infrared images recorded during the active diagnostic experiment are analysed using a variety of statistical texture analysis methods. As the result of the analysis, a set of 259 diagnostic features for each of the five regions of interest is obtained for the 840 recorded images. Two cases of the evolutionary computation are applied in order to search for relevant features. In the first case, the algorithms search for an assumed number of features. In the second case, a number of features are selected automatically by the algorithms. Three types of computational strategies are used. The first one uses a single EA in opposition to the second and third where concepts of weak-strong and external-internal EAs are used. The concepts of weak-strong and external-internal are original ideas of the author. The assessment of feature relevance is performed based upon the results of classification. The classification is performed by a neural network classifier in all evolutionary computation test cases. Finally, the results are compared to those using the classical method of feature selection. The comparison shows that the EA made it possible to obtain higher classification performance by use of a smaller subset of selected relevant features.