Research On Vehicle Target Detection System Based On Image Quality Evaluation And Optimization

In recent years,with the development of sensor technology and the autonomous driving industry,intelligent driving assistance technology has developed rapidly.Among them,target detection technology,as one of the three major tasks of automatic driving,can not only effectively provide intelligent driving assistance for drivers,but also serve as the basic support for other tasks of automatic driving.Most of the existing vehicle detection researches propose or improve various algorithms and models based on open source datasets to improve the accuracy and speed of detection.However,the image quality in the datasets used by these mainstream vehicle detection algorithms is generally better.In actual detection,once the test image has low-quality distortion problems such as low brightness,blurring,and high noise,it is easy to see the detection performance and accuracy drop significantly or even false detection.Aiming at the problems encountered in the detection of low-quality images by the above-mentioned vehicle detection technology,this subject proposes a vehicle detection system based on image quality evaluation and optimization to improve the detection accuracy of road vehicles in complex environments.The high-quality image is then processed by the image optimization technology to enhance the brightness,color,clarity and other dimensions of the low-quality image,thereby improving the performance and accuracy of the vehicle detection algorithm.The main work of this paper is as follows:1)The influence of image quality on vehicle detection accuracy is explored.The most representative target detection network-Yolov5 is selected to train on the existing autonomous driving data set,and the low-quality images of different degrees are detected according to the parameter model obtained by training,and the influence of image quality on the detection accuracy is verified.2)Screening out low-quality images is the basis for subsequent image optimization and vehicle detection.In order to locate the images that need to be optimized quickly and accurately,this paper constructs a quality evaluation method based on low-illumination road scene images.By introducing multi-task learning,a deep neural network model is built,and the two sub-tasks of distortion type recognition and quality score prediction are combined to learn regression and classification tasks;at the same time,a new image quality is produced based on the automatic driving data set based on the sorting principle.The evaluation data set is used for the training of the quality evaluation model of the automatic driving scene;and the network is verified and analyzed on the basis of the data set and the collected real road scene images,so as to accurately evaluate the image quality and realize the screening of low-quality images.3)Based on the output results of image quality evaluation,an optimization method based on low-illumination road scene images is constructed.By introducing deep learning into Retinex theory to improve it,the powerful learning ability of convolutional neural network is used to replace the traditional method to estimate the illumination component in Retinex theory,so as to realize the construction of image optimization model;Image optimization,an image optimization dataset specially designed for vehicle detection tasks is made.The dataset consists of 7480 original data images containing road scenes and their corresponding low-quality images,which are used for the training and verification of image optimization algorithms;use This dataset trains the network,and restores the brightness and color of low-quality images according to the training model,so as to obtain clearer and more natural images.4)Carry out algorithm transplantation and performance analysis after transplantation of the constructed target detection system based on image quality evaluation and optimization.In the performance analysis,corresponding experimental methods are designed to verify.The final experimental results show that the vehicle detection system based on image quality evaluation and optimization proposed in this paper can effectively identify low-quality images and optimize them.The accuracy of recognition success is about 6%higher than that of low-quality images.At the same time,it is verified that the vehicle detection system based on image quality evaluation and optimization proposed in this paper can provide better performance for vehicle detection in lowillumination road scene images.