Hyperspectral Image Ship Detection Based Upon Two-Channel Convolutional Neural Network and Transfer Learning

摘要

Ship detection has been a research hotspot in hyperspectral image target detection. Traditional algorithms are limited by their low detection rate. Convolutional neural network, as a common algorithm in deep learning, has an outstanding performance in features extraction and has been widely discussed in target detection, including ship detection. This paper proposed a novel two-channel convolutional neural network model which is composed of spectral and spatial channels. We picked 1D spectrum vectors from hyperspectral images as the input for spectral channel to extract spectral features. The spatial features were extracted using 2D spatial images by spatial channel. Meanwhile, transfer learning is applied to solve the limitation of 2D spatial images. Experimental results on Hyperion data demonstrate that the proposed algorithm upgrades ship detection rate effectively while suppressing false alarm rate.