Research On Autofocus Algorithm Based On Machine Learning

Autofocus systems are currently widely used in industrial measurement,microscopy,and aerospace.Common auto-focusing systems are divided into active and passive types.This article focuses on passive auto-focusing systems.The performance of the passive autofocus system is determined by the autofocus algorithm.The autofocus algorithm currently has some shortcomings,such as poor noise immunity of the evaluation function,unreasonable window construction,and the focus search algorithm cannot find the accurate focus.The degree is not high,and the real-time performance is not good.As a cross-disciplinary method,machine learning technology has good accuracy in solving classification and regression problems.This paper mainly studies the sharpness evaluation function and focus search algorithm in the autofocus algorithm,combines machine learning technology and focus search algorithm,and proposes a focus search algorithm based on decision tree.Using the learned decision tree to represent the step selection process can effectively reduce the number of searches during the focus search process and improve the real-time performance of the algorithm.The main work of this article is as follows:(1)A variety of definition evaluation functions were selected.Based on the proposed qualitative and quantitative indicators,based on several sets of image sequences collected,a comparative experiment was conducted to select the definition evaluation function with the best overall performance.(2)Collecting a large number of image sequences for feature construction and labeling,using feature variance values for feature selection,and generating the final training data,and using it to learn the step size selection algorithm.(3)The proposed focus search algorithm based on decision tree is compared with the traditional Fibonacci algorithm and global search algorithm on 5 sets of test samples.The results show that the average search times are reduced by 54.8%,and most of the peaks can be searched.The proposed algorithm has good performance in real-time and accuracy.