Research On Infrared Small And Dim Target Detection And Active Tracking Based On Deep Learning

Infrared dim and small target detection and active tracking is the core technology of infrared search and track(IRST),which plays an important role in civil aviation safety monitoring.Due to the extremely long imaging distance and the imaging characteristics of infrared detector itself,infrared dim small target always presents the characteristics of few imaging pixels and extremely low local signal to clutter ratio in the image,which brings great difficulties to the detection and tracking of infrared dim small target,resulting in the situation of low detection accuracy and high miss detection.In this paper,based on the previous work,in order to improve the detection accuracy and recall rate of sky background with extremely low local signal to clutter ratio,aiming at the imaging characteristics of medium wave infrared telescope and infrared dim small target,from the angle of energy convergence and infrared dim small target detection algorithm itself,the automatic focusing algorithm and infrared dim small target detection algorithm of infrared telescope are improved The main research and achievements include the following two points:(1)This paper analyzes the image characteristics of infrared dim small target in the state of quasi focus and out of focus of Cassegrain medium wave infrared telescope,and obtains the importance of energy convergence for infrared dim small target detection.The infrared telescope system has the advantages of long imaging distance,short depth of field,more severe image blur caused by defocus,and the image formed by the telescope is constantly changing due to atmospheric refraction,resulting in the low success rate and efficiency of traditional focusing algorithm.In order to improve the success rate and speed of auto focus,a hill-climbing method with variable step size is adopted.The accuracy of definition evaluation is ensured by using the method of getting the median of image definition for many times.The hill-climbing method of driving amount and acceleration is used to reduce the instability in the focusing process and the number of steps required in the coarse focusing process.The algorithm has been applied in the actual MWIR telescope system.The experimental results show that the number of focusing steps required by the algorithm in the coarse focusing stage is reduced by 12.8% compared with the traditional mountain climbing method,which meets the needs of the infrared telescope system.(2)Aiming at the imaging characteristics of infrared dim small target,such as less pixels and low signal-to-noise ratio,a convolution neural network based on Yolo and traditional image processing algorithm is proposed.Four consecutive images are used,and each image is filtered and inserted into the image sequence to form a total of eight images as network input.After convolution network operation,the position information of the target is output.The experimental results show that,compared with single frame input,the network with image sequence as input has high recognition rate and low false alarm rate;at the same time,compared with pure image sequence input,the network with filtered input can greatly improve the recognition of static or slow-moving small targets,especially suitable for infrared tracking equipment with poor imaging quality.Finally,the whole control system is deployed to FPGA + NVIDIA TX2 platform to detect and track the infrared dim target.On the basis of Cassegrain reflective medium wave infrared telescope,this paper introduces the relevant theoretical knowledge of infrared dim small target detection and tracking based on deep learning in detail,and further studies the shortcomings and difficulties in the current research work,puts forward the auto focusing algorithm suitable for telescope system,and the dim small target detection model,and solves the problem This paper solves the problem of infrared dim small target detection under extremely low local signal to clutter ratio,and compares with the classical model.Experiments are carried out on the hardware equipment to verify the effectiveness of the proposed algorithm and model,which provides some experience for future research.