Research On Key Technologies Of Target Detection Via Infrared Line Array Sensor System

With the continuous development of infrared imaging technology,infrared imaging system has the advantages of long detection distance,high detection sensitivity,strong anti-jamming ability,all-weather operation and so on.It plays an important role in the field of industry,security and national defense.In infrared imaging system,infrared target detection and recognition algorithm plays an important role.In recent years,China has spent a lot of effect on processors.However,the types and performance of these porcessors are still far behind the world's advanced level.Therefore,based on the existing platform,focusing on the objective difficulties of infrared target detection in complex background,this paper proposes an extensible heterogeneous computing framework,and designs the corresponding infrared target algorithm.This paper focuses on the following aspects.(1)Research on infrared target pre-detection algorithm optimized for line scan detector with high speed.The algorithm defines a discrete roundness function according to the band noise characteristics of line array detector,and proposes a fast classification method of noise and target by using the roundness of suspected target,and finally obtains the target by local segmentation.The computational speed of this method has excellent scalability.It can provide 40 million pixels per second target recognition performance on domestic DSP,and has ultra-low false alarm rate and good recognition rate.(2)Research on infrared small target detection algorithm based on gradient enhancement.Aiming at the scene which is difficult to detect in(1),this algorithm proposes a multi-scale fusion enhancement algorithm for infrared target in gradient space by analysing the gradient symmetry of infrared target and the shortcomings of linear detector correction method.Firstly,the forward difference map of the image is calculated,the complementary pixels in the difference map are found,the contrast of the complementary pixels is enhanced,and the image is restored by integration.Then,the enhanced restored images of different scales and directions are fused and superimposed.Finally,the adaptive segmentation threshold is calculated by using the peak to peak value of background clutter to segment small infrared targets.Experimental results show that the algorithm has good small target detection ability,excellent recognition rate,low computational overhead and false alarm rate in complex scenes with ultra-high contrast.(3)Infrared target detection difficulty estimation algorithm based on deep learning.It is difficult to use deep learning to detect infrared small target end-to-end directly because of lack of training set,while traditional feature engineering target detection algorithm still has room to improve the recognition rate of infrared small target through reasonable threshold or parameter setting.Therefore,this paper proposes a multi network threshold estimation framework named iRCNN,which uses multiple subnets to estimate the threshold of the detected area and obtain the prediction threshold by weighting-sum,and uses the distribution characteristics of the prediction results of subnets to derive the credibility of the prediction threshold.Finally,the two are integrated to guide the traditional algorithm to determine the threshold.Experimental results show that compared with Ranking-CNN,iRCNN is more suitable for scene-based dynamic estimation of algorithm threshold,and the effect is better than the fixed threshold or super parameters set by human.(4)Low frame rate track matching algorithm based on motion characteristic modeling.Due to the low frame rate of the image produced by the line array infrared detector and the poor correlation between the small target frames,the traditional multi frame infrared small target detection algorithm can not estimate the track well.Therefore,this paper models the motion characteristics of the infrared small target,estimates the maneuverability of the target by using the historical motion characteristics,then determines the search space,and finally limits the searching area to discard the suspected infrared small target adaptively,then the track is obtained.Experiments show that the method has good detection ability for both low speed and high speed targets.(5)Design of a real-time target detection system.In view of our country's pursuit of independent chip production,this paper proposes a heterogeneous computing framework by using the various technologies mentioned in this paper,which integrates the advantages of DSP and general-purpose computer to realize real-time target detection.Firstly,before the operation of the system,the iRCNN network is used to pre calibrate the background threshold of the whole airspace,so we can reduce the performance requirements of deep learning method;secondly,the real-time processing ability of DSP is used to detect the whole infrared image scene in real time,and the corresponding coarse detection area is generated;then,the gradient space small target detection algorithm and the predictive threshold of iRCNN are combined to detect the target from the coarse detection area,finally,the suspected target is segmented,and the track matching algorithm is used to integrate the target output results.Under the premise of providing high-performance target detection capability,the framework reduces the process delay by fully optimizing the algorithm flow,which has high practical value.