Remote Sensing Images Target Recognition Based On Deep Convolution Neural Network

In recent years,Target recognition technology has also achieved leapfrog development in applications such as face recognition,target tracking,autopilot,defense security,and remote sensing image analysis with the rapid development of artificial intelligence,high-performance computer hardware and cloud computing,and target recognition algorithms proposed in different fields also exhibit superior performance that competing with human capabilities.Deep learning algorithms have been widely introduced into remote sensing applications as a hot topic in machine learning.Remote sensing target recognition algorithms face enormous challenges due to the huge amount of calculations caused by the massive data generated by hundreds of remote sensing sensors every day and the complexity of the surface environment seriously affect the efficiency and accuracy of data analysis algorithms.The two main difficulties are how to improve the real-time processing performance of the algorithm and diversified feature extraction and abstract information expression overcoming complex backgrounds.Because the deep learning algorithm can extract representative and discriminative abstract features of typical features from the image in a multi-level learning method for remote sensing big data analysis.Therefore,aiming at the problems encountered in remote sensing target recognition,two kinds of deep convolutional neural network algorithms based on regional suggestions and regression-based learning are introduced,which are applied to identify the shared data sets of remote sensing targets after data amplification,exploring the data processing mechanism and convolution features representation of deep learning algorithms.The characteristics and performance of two types of deep learning algorithms are compared and analyzed.The main content of this article includes:(1)Amplify remote sensing dataset for target recognition.Building rich,high-quality training datasets is the key to train deep learning algorithm models,and high-quality dataset can contain more comprehensive multi-scale and multi-angle observation infonnation of different interested targets in different environments.In turn,the overall accuracy of the parameters and target recognition of the deep learning algorithm model can be optimized.Due to the influence of lighting conditions,terrain and atmospheric environment,the target’s expression form in remote sensing images is very complex,which directly affects the effect of target recognition algorithms,especially for small-scale targets.To solve this problem,based on the existing shared data sets,this paper constructs an remote sensing image sample set with more imaging conditions by using artificial markers,which provides a more solid foundation for subsequent target recognition algorithm research,The dataset contains eleven types of remote sensing image targets:aircraft oiltank,overpass,ship,baseball diamond,tennis court,basketball court,playground,harbor,bridge and vehicle.(2)Research on remote sensing images target recognition algorithm based on deep learning.This article compares the performance of two types of deep convolutional neural networks:Faster Region Convolutional Neural Network model based on regional proposal and SSD model based on regression learning,applied to multi-class remote sensing target recognition.The recognition results show that,in terms of accuracy,the average accuracy of the SSD is 81.3%,slightly higher than 73.2%of the Faster R-CNN.However,the accuracy of SSDs for small-scale targets,such as cars,is only 54.7%,which is much lower than that of Faster R-CNN.In terms of efficiency,the frame rate of SSD processing Jilin-1 video star data is as high as 46 frames,far higher than the 7 frames of the Faster R-CNN,demonstrating excellent real-time video target recognition potential.