Research And Application Of Image Enhancement Algorithm In Driving Image Processing

Automobile data recorder is now widely used in all kinds of vehicles,and its driving image has a very wide range of applications,such as parking monitoring,judicial evidence.However,limited by many factors,such as illumination and acquisition equipment,the video images collected from some vehicle recorders in the market are facing some problems.Second,the image acquired during the day is characterized by fuzzy edges and low resolution.The most direct way to improve the night image quality and image resolution is to modify the hardware in the vehicle recorder system,but the cost is high and the manufacturing process is not easy to improve.Therefore,from the point of view of algorithm to achieve low-illumination image enhancement and image super-resolution reconstruction has become a research focus in the fields of image processing and computer vision.In order to obtain the ideal high-resolution image with clear night,sharp edges and fuzzy blocks,this paper focuses on the single image enhancement and super-resolution reconstruction in driving image.The main contents are as follows:(1)For the images collected under low illumination at night,four traditional classical algorithms are firstly used to carry out simulation experiments in this paper.Second in view of the traditional algorithms can’t retain the image detail problem very well,inspired Retinex method,is presented in this paper using the improved convolution of the neural network method of image decomposition,the decomposition of illumination image did histogram equalization processing,do to reflect the image denoising processing,the image contrast and fidelity effect has significant improvement;Considering the convolution of the neural network training need pairs of data sets,it is well known that the same scenario of image data acquisition faces challenges in pairs,so at first,this paper collected 15000 traffic data set and data enhancement,secondly use the unsupervised methods(no need to data sets in pairs)realized with low illumination image enhancement,and from two aspects of subjective and objective to enhance image has carried on the evaluation,evaluation results show that the image quality got obvious improvement.(2)In view of the amplification of image noise taken under low illumination,this paper analyzes the existing denoising technologies.In this paper,the method of bilateral filter denoising is used to carry out the simulation experiment.Secondly,a model based on convolutional neural network was built to process the image noise.Simulation experiments were carried out with the algorithm and its performance was evaluated.The simulation results showed that compared with the original noise image processed by the algorithm,the peak signal-to-noise ratio increased by 4.15 db and the structural similarity increased by 28%.(3)In view of the features of the collected image,such as fuzzy edges,local block blur and low resolution,this paper studies the reversibility of image sampling,and conducts a simulation experiment with Laplacian method.In view of its limited effect,the super-resolution reconstruction of a single low-resolution image is carried out by means of generating antagonistic network.In the end,the problem of incoherence and inlifelike between frames in video is studied,and the super resolution of video is reconstructed by using the generation antagonism network.