A Method for Extracting and Selecting Features of Thermal Infrared Remote Sensing Images of Harbor Targets

Abstract

To improve the inadequate effectiveness of thermal infrared (TIR) image recognition due to low contrast, striping noise and low spatial resolution, this paper conducted research with respect to harbor targets. It adopted sample features like texture, geometry, shape and statistics, selected the best combination of image features with an attribute selection evaluator, and chose the best classifier and the best classifier parameters according to the accuracy of recognition. This process can generate the best features for the classification of TIR images with a certain degree of robustness. The results of the experiment indicate that 1) the libSVM (a support vector machine), after parameter optimization, is the most accurate classifier and the best result can reach 82.4137%; 2) the one-dimensional statistical features of pixel value, mean value, variance, mid-value, and gray level histogram are more meaningful for the recognition; and 3) daytime images can be more accurately classified than nighttime ones.