Feature extraction, selection and classification code for power line scene recognition

Abstract

Detection and avoiding of power lines and cables is a critical issue in aircraft flight safety. Despite various improvements in image analysis literature, most of the safety issues depend on visual capabilities of pilots. It is aimed that proper scene detection methods may help the pilot by igniting alarms. The presented work basically considers frequency based features (in the real valued discrete cosine transform — DCT domain) as candidates of signatures for existence of power lines in the image. Since DCT provides spectral distribution along all frequencies, a domain-search method is adopted to see where in DCT samples the most signatures are carried. The developed software searches most candidates of DCT regions, compares them with performances of other saliency-based popular methods (such as LBP and HOG), and tests their representation powers via various classifiers. Image pre-processing and feature extraction parts are implemented in MATLAB™ R2013b simulation environment, the classification step was implemented on WEKA 3.8.0. A flowchart is formed where pre-processing is sequentially performed, and features are simultaneously extracted; finally, the outputs are fed to WEKA environment for classification evaluation.