Research On Satellite Image On-board Target Recognition Based On Local Invariant Feature

With the development of space remote sensing technology and the continuous improvement of satellite image resolution,an increasing number of high-resolution satellite images apply to the civil and military fields.Especially in military fields,satellite image recognition is an important way to obtain military information.Since the on-board process ability of remote sensing satellite in China is poor for time being,recognition and processing work of satellite image is carried out mainly on the ground.Due to the ability of satellite data transmission and the distribution of ground stations,image acquisition process is very complex.Therefore,it takes data about one day for transmission from acquisition to the ground station.However,on-board target recognition can process satellite images on orbit and transmit the processed data to the ground command center through relay satellite or telemetry system in time.In this way,the acquisition time is shortened greatly,which ensures the timeliness of the remote sensing information.In this dissertation,an on-board target recognition method for satellite imagery is proposed.According to the characteristics of remote sensing images and the problems existing in the on-board processing,this method is proposed to be combined with the development trend of the target recognition technology.Deep researches with the step of target recognition,such as feature detection,feature description and match recognition,are also made.In addition,an on-board fast target recognition system based on the characteristic of proposed method is designed.The system will provide theoretical basis and suggestions for practical engineering application.Aiming at ensuring the speed and accuracy of the target recognition,a fast feature point detection algorithm based on the scale space theory is proposed.The algorithm builds the scale space by down sampling the images and detects the feature points with AGAST algorithm in every scale space.However,the scale space will lead the feature point to offset.Considering this problem,a method based on the feature point local relocation to location of feature points is proposed.Experiment results show that the algorithm is robust to scale,viewpoint and illumination transform.What's more,the algorithm uses a shorter time,and its precision of feature points detect is higher.In view of the complex and changeable characteristics of remote sensing imagery,especially the obvious changes of illumination,a FREAK description algorithm based on feature point intensity normalization is proposed.First,by adjusting the FREAK sampling model,sampling points are made to focus on the edge and contour feature.Then,the dominant orientation of the descriptor is estimated by sampling point's gradient.At last,the pixel intensity is normalized use of the neighborhood of feature point.Experiment results show that the performance of proposed algorithm is better than the others' in terms of scale,rotation and illumination transform and the performance of viewpoint transform satisfies the on-board process.In addition,the algorithm uses a relatively short time which is suitable for fast processing of remote sensing images.With the large amount of satellite image data and the low processing ability on the satellite,a hybrid spill tree data search method for satellite image on-board target recognition is proposed.The algorithm redundant segments the data in feature preprocessing stage.By giving up the edge of the sparse data points,the algorithm adopts segmentation method based on center point.In feature matching stage,the algorithm measures the distance between feature vectors by using XOR operation.SIFT feature matching strategy is also improved by using K nearest neighbor distance instead of second nearest neighbor distance.Finally,RANSAC algorithm is used to eliminate false matching points.Experiment results show the proposed search algorithm to recognize the target in remote sensing image has higher efficiency,shorter processing time and higher matching accuracy.When the solar altitude is low,there are many shadow areas on the satellite images,which leads to a low target recognition rate.Within view of this problem,a target matching recognition method based on super pixel segmentation and reconstruction is proposed.First,the orientation of aircraft needs to be estimated with histograms of oriented gradients.Then,an improved SLIC super pixel segmentation algorithm is provided.By comparing texture feature instead of color feature space,the pixels with the same features are clustered.Lastly,the super pixels are reconstructed through target template images and orientation.And targets are recognized by feature matching.The experiment results show that the algorithm is more accurate with target segmentation and wipes off the shadow areas more efficiently.Compared with the situation of superpixel segmentation before,the method in this dissertation is more precise.In engineering application,an on-board rapid target recognition system is developed based on FPGA+DSP integreted platform.The ideas of OpenVPX bus design and dual DSP parallel pipelining are adopted to ensure versatility and scalability of the system.In this way,the processing efficiency of the algorithm can be improved and foundation for the on-board rapid target recognition can be laid.In addition,high speed interface provides a strong guarantee for the transmission of large amount of remote sensing image data on satellite.At last,the result of target recognition based on original satellite image data is shown,which verifies the feasibility of on-board target recognition.