|Table of Contents|

Remote sensing scene recognition using unsymmetrical nonlocal convolutional neural network(PDF)

Journal of Forestry Engineering[ISSN:2096-1359/CN:32-1862/S]

2020 No.06
Research Field:
装备与信息化 执行主编:周宏平
Publishing date:


Remote sensing scene recognition using unsymmetrical nonlocal convolutional neural network
XU Feng SUN Wanyan
College of Information Sciences and Technology, Nanjing Forestry University, Nanjing 210037, China
convolutional neural network remote sensing scene image recognition intelligent forestry
Compared with traditional shallow learning methods, deep learning based methods have the advantages of high efficiency, outstanding performance and strong expression ability. Among these methods, the image processing technology based on the convolutional neural network is widely used in the fields of image classification, semantic segmentation and object detection. In particular, Nonlocal convolutional neural network (Nonlocal CNN) is an enhanced deep convolutional neural network model, which has attracted much attention in recent years. In image processing tasks, the Nonlocal CNN can be used to obtain the global contextual rather than the limited local information of an input image, which is proven to be beneficial for image recognition tasks. In a large number of computer vision tasks, the effectiveness and feasibility of the Nonlocal CNN have been proved. Considering the fact that there are numerous repeated pixelpatches and objects in some categories of remote sensing images (i.e., forest, farmland and beach), the repeated calculations for the same pixelpatches not only would reduce the model efficiency but also may worsen the performance of the model. To address this problem, this study proposed an unsymmetrical nonlocal convolution (UNC) module to reduce the calculation time and space complexity of the Nonlocal CNN, so that the model efficiency can be improved. Specifically, a downsampling operation on feature maps of Keyvalue pairs in the classical nonlocal convolutional module was performed firstly; then, a global object context estimation and weighted feature maps of the input image was obtained by the pixelwise multiplication of Query and Keyvalue pair; finally, the residual connection and weighted feature maps were added together as the output for the UNC module. Compared with the classical Nonlocal CNN, the UNC has the same implementation process as the Nonlocal CNN, but the proposed UNC was an efficient and general context acquisition module that can be used in a variety of deep learning CNN models. In order to demonstrate its effectiveness, experiments were carried out on four public remote sensing object recognition datasets. The experimental results showed that UNC could reduce the model space complexity by up to 24.53% as well as 49.1% on the floating point operations, and obtain the recognition accuracies of 96.63%, 99.16%, 98.90% and 96.28% in RSSCN7, UCML, WHURS19, and AID, respectively, which demonstrated that UNC had a strong practicality and universality.



Last Update: 2020-11-11