Design Of Signal Processing Module For Infrared Focal Plane Detector Based On Machine Learning

With the rapid development of infrared imaging technology,infrared thermal imaging systems have become more and more widely used in military and civilian applications.The infrared focal plane array detector is the most important detector device in the infrared imaging system.The signal processing module based on it is also rapidly developing in the direction of high performance and intelligence.Based on this,this paper comprehensively designs the signal processing module of infrared focal plane array detector based on machine learning algorithm.The hardware architecture,infrared image non-uniformity correction and infrared image real-time target detection are the three main parts of the module.High-performance hardware system design,effective infrared image non-uniformity correction and real-time accurate infrared image target recognition and positioning.In terms of hardware infrastructure,based on system design requirements,this paper builds a high-performance hardware platform that uses DSP(Digital Signal Processing)and FPGA(Field-Programmable Gate Array)as the processing core.The system uses the FPGA,the DSP board and the interface board to constitute the three major components of the hardware.Among them,FPGA is used as the main logic generation chip,which controls the detector's image acquisition,image preprocessing and image stream transmission.DSP uses powerful parallel processing capabilities to calculate and adjust infrared image information and run key image processing algorithms.The interface board is designed with a wealth of interface resources to enable interoperability between hardware and detectors,computers,and imaging devices.At the same time,this article addresses the communication and data read and write issues between FPGAs and DSPs,and designs and implements the SRIO(Serial Rapid I/O)high-speed data interconnect function,and introduces the DDR3(Double Data Rate 3)external memory control module to make the system Performance further improved.The area of the largest single board in the overall hardware system is 90mm*50mm,and the entire module height is controlled within 50 mm,achieving a miniaturized design goal.Through further imaging experiments,the system has completed the collection,processing and transmission of infrared images,meeting the requirements of image acquisition and processing rate,and providing clear imaging.Because the infrared focal plane array imaging tends to have obvious non-uniformity characteristics,this paper has thoroughly studied the traditional calibration-based and scene-based infrared image non-uniformity correction algorithms,and analyzed the traditional algorithms' disadvantages in the face of detector nonlinear response,parameter drift,edge texture blur,etc.On this basis,this paper selects adaptive backpropagation neural network non-uniformity correction algorithm as the infrared image non-uniformity correction scheme,and introduces image gradient information to the hidden layer,loss function and iterative step of the traditional network respectively.This paper realizes the removal of high-frequency noise in infrared images while preserving image detail information and accelerating the convergence speed of the algorithm.In the simulation experiment,the correction effect of the improved algorithm is more stable than the traditional algorithm,and the signal PSNR is higher at the same convergence speed.After the hardware system is implemented,the infrared continuous imaging non-uniformity correction effect is good.Based on the two-point correction,the improved algorithm successfully removes the noise and edge noise in the infrared image and at the same time protects the image texture details and meets the system design requirements.The target recognition and positioning function of infrared thermal imaging system has a wide range of applications in missile guidance,target detection and vehicle navigation,and it is also a key technology of intelligent infrared image signal processing module.This paper studies the progress of convolutional neural networks in the recognition and localization of visible light targets,and introduces the Faster-Rcnn algorithm in convolutional neural networks to target recognition of infrared images.In this paper,the infrared images collected in the field are divided into training set and test set.The training set is marked and the test set is used to identify and locate the characters and vehicle targets.The final experimental recognition accuracy rate reaches 85%,and the infrared image target can be accurately identified.At the same time,single-frame image detection speed is 0.66 s,which basically meets the requirements of real-time monitoring of targets.The research results of this paper have achieved the system design goals of the signal processing module of high-performance and intelligent infrared focal plane detectors,laying a good foundation for further research on signal processing modules with higher integration and better real-time performance.