Aircraft Detection by Deep Belief Nets

Aircraft detection is a difficult task in high-resolution remote sensing images, due to the variable sizes, colors, orientations and complex backgrounds. In this paper, an effective aircraft detection method is proposed which exactly locates the object by outputting its geometric center, orientation, position. To reduce the influence of background, multi-images including gradient image and gray thresholding images of the object were input to a Deep Belief Net (DBN), which was pre-trained first to learn features and later fine-tuned by back-propagation to yield a robust detector. Experimental results show that DBNs can detecte the tiny blurred aircrafts correctly in many difficult airport images, DBNs outperform the traditional Feature+Classifier methods in robustness and accuracy, and the multi-images help improve the detection precision of DBN than using only single-image.