Research On Image Fast Focusing Technology Based On Convolutional Neural Network

As an effective means of monitoring targets,optical measurement systems have always been valued by everyone.Since the focusing technology used by traditional optical measurement equipment in the shooting range generally needs to rely on additional auxiliary equipment to guide the focusing operation,it does not meet the development requirements of the integration of optical measurement equipment.The image-based automatic focusing system can complete focusing only by relying on image information,and the hardware structure is highly integrated and the calculation speed is fast.It has gradually become the main research direction in the field of automatic focusing.However,the automatic focusing method based on image processing is a method to gradually find the optimal value of focusing.In a real-time focusing system,the slow focusing speed is the primary problem to be solved by this method.On the other hand,convolutional neural networks have achieved outstanding results in the field of image quality evaluation without reference in recent years.The clarity of images can be intuitively evaluated through the network model,with fast calculation speed(no iterative process)and high accuracy(no local optimal problem)and other advantages,it is expected to solve the problem of slow focusing speed.In view of the above background,this paper studies the image fast focusing system based on convolutional neural network.First of all,aiming at the problem of insufficient precision of SMD function in the fine focusing process,an improved SMD evaluation function based on pixel difference is proposed.Based on the original SMD function,this function supplements the comparison of two additional pixels in the horizontal and vertical directions.In addition,it adds the contrast of the target pixel at 45 ° and 135 °diagonally.Through Matlab simulation and auto-focusing experiments,Experimental results show that the sensitivity and noise immunity of the evaluation function have been improved,which is beneficial to improve the imaging accuracy of the focusing system.Secondly,since the image evaluation value cannot directly reflect whether the image is clear or not,an image evaluation network is proposed.By improving the VGG convolutional neural network,the task of recognizing the degree of image blur is realized,and it is compared with the commonly used image recognition network methods.The results showed:The improved VGG network has a recognition accuracy of 97.45%,95.28%,93.65% and 83.54% for the data set containing 4,7,10 and 15 categories of different degrees of ambiguity,respectively,it is proved that the image evaluation network can identify the blur degree of the image well.Then,in view of the problem that the traditional hill-climbing search algorithm large-step focusing easily causes the motor to reverse and affects the focusing accuracy;the small-step focusing increases the number of focusing and affects the focusing efficiency.An improved hill-climbing search algorithm with image evaluation network as the main and image evaluation function as the auxiliary group is proposed.First,the image is identified by the image evaluation network to identify the degree of blur,and then the precise search step is formulated according to the degree of image blur.The image evaluation function is used to control whether the focus search direction is correct.The curve-fitting search algorithm is adopted at the position where the system defocus is relatively low,and the advantages of each algorithm are fully utilized to improve the search efficiency of focusing and effectively avoid the local extreme value during the focusing process.Finally,the relevant algorithms proposed in this paper are applied to the automatic focusing system based on surveillance cameras.The experimental results show that the search speed of the search algorithm in this paper is 2-3 times that of the traditional hill-climbing method,the focusing efficiency is significantly improved,and the number of motor drives is stabilized at 5-8 times,and is not affected by the initial state of the system.The feasibility of the automatic focusing scheme is proved,and it provides a reference for the application of convolutional neural network in the automatic focusing system in the future.