Research On Automatic Focus Technology Of Optical System Based On Image Information

Photoelectric theodolite is a large-scale optical detection device commonly used in modern test range.It is mainly used to automatically capture and track long-distance flying targets that enter the field of view to obtain its flight posture,geometry and flight trajectory.These data are of great significance to scientific research and military defense.Because the measured target is expensive,irrecyclable and extremely fast,if the imaging system fails to process the defocused information in a timely manner,it will result in blurred imaging and a large amount of detailed information will be lost.Failure to collect relatively accurate test data will greatly increase the test cost.At present,photoelectric theodolites mostly use automatic acquisition of object distance,combined with environmental factor compensation to automatically calculate focus.However,because active focusing requires auxiliary equipment to emit signals for detection,it is easy to be anti-reconnaissance when carrying out certain tasks that require concealment,and the arrangement of auxiliary equipment in advance for the venue does not conform to the trend of flexible and portable photoelectric theodolites.Therefore,passive auto-focusing,which does not require the arrangement of auxiliary equipment and does not require the measurement of transmitted signals,has become a popular research direction in this field.The automatic focusing technology based on image information is one of the hot spots in the passive automatic focusing field.This article focuses on research hotspots,closely follows the trend of field intelligence,combined with statistical machine learning methods,for fixed-focus optical systems,training using 2-3 frames of ordered images can predict the multi-classification of the current object distance quasi-focus position model.In this regard,this article mainly carried out the following research work:1)Investigated the important role,development status and development direction of the photoelectric theodolite and its automatic focusing technology in modern military and scientific research,and understood the urgency of further research on the automatic focusing system.And researched and analyzed various kinds of automatic focusing technology at domestic and international2)Analyzed and studied the geometric model of optical system imaging and defocus.Analyzed the main factors that affect the clarity of system imaging. Starting from the principle,the geometrical analysis and derivation of the influence of the object distance were discussed,and the relationship between the focal depth,depth of field and imaging system parameters was discussed.3)This paper analyzed the existing image clarity quantification methods,and selected the power spectrum clarity evaluation function that could meet the needs of focusing.According to the needs of application scenarios,the size of the image and the average brightness were used to design the coefficients of the evaluation function.While reducing noise,it reduced the influence of th uneven image size and uneven brightness on the sharpness evaluation curve.4)According to the general steps and methods of statistical machine learning to solve the problem,the input-output relationship of the auto-focusing problem was analyzed.In order to structure the image data,the power spectrum evaluation function was used to quantify and extracted the clarity of the image, and the calibrated image acquisition position,the ratio relationship between the image quantization value(DFM value)and the difference relationship were used to build some more stable feature,two data sets capable of describing the relationship between two-point values and three-point values was produced.5)Based on the structured data sets constructed in Chapter 4,for the non-numerical discrete data multi-classification problem of image auto-focusing,an extreme random forest algorithm that meets the performance requirements was used for training and modeling.Single decision tree and random forest algorithm were used as the comparison items of extreme random forest algorithm,the main parameters were fixed,and the performance of the algorithm was intuitively evaluated.Through the precision,recall,f1＿scores,and the results displayed by the confusion matrix: Both the two-point value relation data set and the three-point value relation data set could get good classification results through training,but the model obtained by the three-point value data set training was more accurate and stable,and the score was higher.In the three-point data set,the classification model trained by the extreme random forest algorithm ensured that f1＿scores was greater than 0.76,and the precision of single-class prediction can reach 0.92.