Dense Convolution Siamese Network for Hyperspectral Image Target Detection

Compared with grayscale and RGB images, hyperspectral image (HSI) can provide both spatial and spectral information of ground targets, which makes it possible to improve the efficiency and accuracy of target detection. Therefore, the research of HSI target detection algorithms has attracted widespread concern in recent years. With the development of hardware devices and the arrival of big data era, deep learning algorithms have been successfully applied to image processing, text recognition and other fields. However, due to the complex gathering environment of HSI, it is so difficult to obtain a large number of labeled samples, which limits the application of deep learning algorithms in HSI target detection. Therefore, a dense convolution Siamese network (DCSN) is proposed for HSI target detection, which improves the accuracy in the scenery of small-scale training samples. The main contributions of this paper include the following three points. First, we design a target sample generation method based on improved autoencoder to enhance target training data. Then, a background selection method based on density estimation is presented, which can acquire typical background samples effectively. Finally, a spectral feature extraction method based on dense convolution is proposed to extract the more discriminative spectral features. The experimental results of HSI target detection on Muufl Gulfport and San Diego datasets indicate that our proposed DCSN is able to achieve superior performance than the existing target detectors.