Identification Of Traffic Signs In The Haze Environment Based On CNN

With the development of technology,intelligent driving systems are getting closer and closer to people’s lives.As an important indication on the road,the accuracy of traffic sign recognition is of paramount importance.In actual road traffic,the collected images are affected by real-time weather,and there are situations where the illumination is too strong or too weak or blurred.At the same time,due to the angle problem,there are also deformations and occlusions,which greatly increase the traffic.The difficulty of sign recognition.This paper mainly studies the identification of traffic signs under hail conditions,and proposes a deconvolution algorithm based on feature multiplexing for convolutional neural networks,and proposes a traffic sign recognition network based on adaptive affine transformation network.In the improvement of the dehazing algorithm,this paper fully understands the basic theory of atmospheric scattering model and restores the dark channel algorithm to improve the understanding of the single image defogging process.By analyzing the limitations of the dark channel algorithm,such as the accumulation of errors under multiple calculations,it is not applicable to large-area sky regions.A deconvolution algorithm based on feature multiplexing is proposed.The algorithm uses the atmospheric scattering model as the theoretical basis to obtain a better image defogging effect by establishing a corresponding relationship between the foggy image and the fog-free image.In this network,the feature reuse unit reuses the features of the previous layer,reducing the number of features required for each convolution operation,and improving the ability to express low-level features by combining low-level features with high-level features.,effectively reducing the dependence on the number of network layers.On the other hand,since the algorithm directly establishes the connection between the foggy image and the fog-free image,the error caused by multiple calculations in the dark channel algorithm is largely avoided,and the distortion phenomenon in the sky region is overcome.The problem that makes it impossible to perform operations such as recognition makes the visual effect particularly outstanding.The standards used in this paper are MPSNR and MSSIM.The former illustrates the peak signal-to-noise ratio of the dehazing image after comparing the original image,and the latter illustrates the subjective perception of the image after de-fogging.Comparing the experimental results of fogging in four different data sets,the algorithm proposed in this paper can effectively defogg the image,and the image after defogging is clear.Aiming at the problem of the common deformation of traffic signs captured in natural scenes,this paper proposes an adaptive affine transformation network.The network obtains the required affine transform coefficients by extracting the features of the input image to achieve the purpose of adaptive affine transformation of the input image.This section of the experiment is divided into two parts: The first part uses the Belgian traffic sign dataset to verify the effectiveness of the proposed adaptive affine transformation network.In this experiment,two different networks were used for experimental comparison,and the accuracy,recall rate,F1 score and Macro-F1 were used to verify the description.The adaptive affine transformation network proposed in this section can be used.Effectively improve the recognition rate and reduce the dependence on the number of layers of convolution;the second part of the experiment is to combine the proven adaptive affine transformation network with the improved VGG network to complete the traffic signs required for this article.Identify the model.In this part of the experiment,the German traffic sign data set GTSRB was used,and the improved VGG network was compared with the improved VGG network with adaptive affine transformation,which was also compared with four different evaluation criteria.Comparing the experimental data,the proposed algorithm has good recognition rate and practicability.