Evolutionary Computing and Transferable Learning based Small Airborne Target Detection and Classification under Stealth Condition
The exponential rise in technologies has broadened the horizon for industries as well as socio-defense vertical to enable more efficient and precise decision making. Defense system being one of the most vital and critical need in contemporary conditions where intruders exploit different approaches to enter unauthorized region demands more efficient surveillance mechanisms. In the last few years, drone technologies or small flying objects including airborne strategic systems have increased significantly. On the other hand, the flexible government policies and innovative work culture has broadened the horizon of drone application for commercial purposes such as video-suiting, food-delivery, medical delivery, transportation etc. Noticeably, being a small target region under highly cluttered and dense environment, employing conventional radar technology seems limited or cost consuming. On the other hand, the development in stealth technology has enabled managing radar cross section minimum, thus making small target detection highly complex. In this paper, we have developed a highly robust and efficient radar signal processing (RSP) model for small moving airborne detection system. Unlike classical RSP methods, our proposed system employs Continuous Wavelet Transform (CWT) which performs Time-Frequency Representation and wavelet image generation for the input reflected signal or radar cross section information. In the sub-sequent phase, to enable precise target detection an evolutionary computing based Fuzzy C-Means Cluttering (FCM) has been developed for moving airborne target segmentation, which gives a precise target representation and its dimensional characterization. To further identify the type of exact target under probable stealth environment we have applied a transferable deep learning method named AlexNet Convolutional Neural Network (AlexNet-CNN), which performs target RCS and wavelet image classification as cylinder and cone, characterizing the precise target types and threat severity. Our proposed AlexNet-CNN based model achieves accuracy of 96%, precision 97.7%, recall 96.8%, and F-Measure of 97.2%.