Research On Improvement Of Target Detection Algorithm For Foggy Road Based On YOLOV3

In winter,the frequency of foggy weather increases due to temperature inversion.Fog scene not only interferes with the driver's subjective judgment ability,but also seriously affects the image recognition performance of automatic driving,thus causing some hidden dangers to the safe travel of vehicles.In order to minimize the influence of fog scene degradation and improve the accuracy of image recognition system,it is necessary to preprocess the original image before target detection,increase the input image information and improve the contrast.At present,there are two different research directions of image defogging algorithm,which are image restoration and image enhancement.The former mainly uses atmospheric scattering model and prior knowledge to respond to defogging.The brightness of the image processed by this method is low and the generalization ability is weak.The latter is mainly through the transformation of image space domain and frequency domain to remove image noise and improve visual effect.As the third version of YOLO series detection algorithm,YOLOV3 is favored by the industry because of its fast detection speed and high accuracy.In the process of detection,by optimizing and improving the structure of the original YOLOV3 network model,the detection and recognition ability of the model for foggy road targets can be further improved.This paper mainly studies the Retinex defogging algorithm based on image enhancement and the optimization and improvement of YOLOV3 neural network model:(1)Aiming at the problems of distortion and halo in the existing image defogging algorithms,an improved single scale Retinex defogging algorithm is proposed.Firstly,according to the transformation of scene depth and fog concentration in the image,the adjustment factor is added to adjust the local image definition through the target position.Then,the bilateral filter is used to replace the Gaussian low-pass filter in the original algorithm for illumination estimation,and the gray value and spatial position information of the original image pixels are retained.Finally,the sigmoid function is used to enhance the reflection image to obtain the final defogging image.Through the subjective and objective evaluation of the image,the defogging effect is improved.(2)In order to meet the test requirements of image defogging,the executable program of image defogging is compiled based on Opencv in Visual Studio development environment.Through the integration of defogging algorithm to achieve different algorithms of image processing.(3)In order to solve the problems of the average accuracy of YOLOV3 target detection algorithm and the slow forward reasoning speed,an optimized and improved deep learning network model structure is proposed.Firstly,a feature scale is added to the network model,and an integrated SPP spatial pyramid pooling structure is introduced into the convolution layer to recalculate the size of the prior box anchor box,which improves the confidence of target recognition.Then,instead of momentum optimizer,Adam optimizer is used as gradient descent strategy,and batch normalization layer and convolution layer are combined to speed up forward reasoning.Finally,GIOU loss is used as the measure of loss function to improve the performance of network model.