Research On Several Problems In Intelligent Target Recognition Of Multi-source Remote Sensing Image

To promote the rapid development and strategic iteration of the new generation of artificial intelligence technology,many countries worldwide are actively deploying artificial intelligence development strategic plans.Since 2017,China has promoted the development of artificial intelligence with an active and open attitude and successively issued government documents,including the "New Generation Artificial Intelligence Development Plan" and "Guidelines for the Construction of the National New Generation Artificial Intelligence Standard System".The innovation and development of artificial intelligence has become a key national layout area,and artificial intelligence technology will continue to have unique advantages in military and civilian fields.Especially in military monitoring,artificial intelligence technology has considerable strategic application value.In recent years,China’s airborne,spaceborne,and missileborne imaging systems have developed rapidly and have multiplatform,multispectral,multiband,multimode,multipolarization and high-resolution earth observation imaging capabilities.Massive high-resolution multisource remote sensing image data have been accumulated.Compared with the rapid development of imaging technology,the intelligent interpretation of multisource remote sensing images is still in its infancy.The ability to obtain high-value intelligence information from multisource remote sensing images is weak,and it is difficult to meet application requirements in terms of accuracy and real-time performance.Therefore,how to use advanced artificial intelligence technology to achieve efficient and intelligent interpretation of massive high-resolution multisource remote sensing image data,extract high-value intelligence information,and improve the level of intelligence is a key technical problem that urgently needs to be solved.Facing future practical engineering applications,this paper takes multisource remote sensing images as the basic data source,takes remote sensing image target recognition as the task orientation,focusing on data,algorithms and computing power,and then conducts research on multisource remote sensing image intelligent target recognition technology to realize the all-round exploration of intelligent interpretation of multisource remote sensing images.This will provide key technical support for promoting the intelligent transformation of remote sensing data into remote sensing information.The main research contents of the paper are as follows:(1)Most of the existing algorithms are aimed at specific close-range targets and are rarely involved in the target recognition of large-scale remote sensing images of targets such as aircraft,oil tanks,and ships.To this end,an optical remote sensing image target recognition algorithm based on deep kernel learning is proposed.The algorithm adopts a two-stage mode of coarse detection and fine detection.First,a target region extraction algorithm based on the visual saliency mechanism is designed to locate the region of interest;furthermore,a saliency feature analysis and recognition model based on nonlinear deep kernel learning is constructed to achieve target recognition.The effectiveness of the proposed method is verified based on three public and self-built optical remote sensing image target datasets.(2)In complex military application scenarios,the background interference is severe,and the qualities of SAR images with different parameters are very different,which greatly affects the performance of SAR target recognition.To this end,an SAR image target recognition system based on fusion PCA and deep learning is proposed.The system first uses PCA to perform statistical dimension reduction on the input SAR image to retain the significant feature information of the image;furthermore,the extracted dimension-reduced image is input into the Faster RCNN model to achieve accurate detection of SAR targets.In addition,based on the Xian Y-7 aircraft test platform,a set of airborne SAR data acquisition systems is constructed,and a real scene SAR image target dataset is built.Based on the self-built SAR image target dataset,the effectiveness of the proposed method is verified.(3)Due to factors such as high shooting height,rapid changes in viewing angle,variable target size,image distortion,and relative motion between the target and the UAV,the task of UAV video target recognition is extremely challenging,and it is difficult for existing algorithms to balance detection accuracy and speed.To this end,a parallel ensemble lightweight deep learning framework is proposed,which adopts global and local joint detection strategy to achieve real-time detection of multiple targets in UAV videos.The framework combines lightweight deep learning and template matching to fully mine significant feature information and integrates multiprocess and multithreading mechanisms to speed up real-time target detection.Furthermore,from the perspective of engineering applications,a target detection system for UAV ground stations is designed to realize real-time accurate detection of ground video targets and display the detection results stably in real time.Based on the self-built UAV remote sensing image dataset,the effectiveness of the method proposed in this paper is verified.(4)To realize the transformation of UAVs to the direction of autonomous intelligence and meet the balanced requirements of target detection accuracy and real-time performance of airborne platforms with limited hardware resources,this paper develops a deep learning model compression and acceleration technology system for airborne platforms.First,this paper studies the compression and acceleration technology of deep neural networks and designs a compression and acceleration strategy more suitable for the target recognitron framework of multisource remote sensing images by integrating a variety of compression and acceleration methods.Furthermore,an embedded deep learning solution based on DSP+FPGA is designed,with DSP as the main processor to realize the main process control and FPGA as the coprocessor to realize AI acceleration of deep neural networks to meet practical application requirements.Based on the self-built UAV remote sensing image dataset and demonstration platform,the effectiveness of the method proposed in this paper is verified.The research work in this paper can provide key technical support and engineering application references for multisource remote sensing image target recognition.It can realize rapid and effective auxiliary decision-making for reconnaissance images,improve the utilization rate of image reconnaissance equipment,strengthen the combat effectiveness of image reconnaissance equipment,and enhance the timeliness and intelligence level of battlefield intelligence support.