Research On Traffic Sign Image Recognition Algorithm Based On Haze Weather

Traffic sign recognition is one of the key links of intelligent transportation system and driverless technology.The recognition accuracy directly affects the safety and reliability of intelligent transportation systems and driverless vehicles.So the traffic sign recognition algorithm based on haze weather has very important research value and practical significance.Combined with the project requirements,this paper focuses on two aspects for in-depth research: defogging image processing of traffic signs with fog and traffic sign recognition algorithm based on CNN,and completed the simulation experiment of the algorithm.The main research of this dissertation is as follows:(1)In order to solve the halo effect of traditional dark primary color estimation algorithm,an adaptive dark primary color estimation algorithm with multi-scale windows is designed.it is obtained by edge detection of the image and changing the window size according to the depth of field edge information.In order to solve the distortion of bright area after defogging of traffic sign image,an image defogging algorithm based on adaptive transmittance restoration is designed.Firstly,the atmospheric dissipation function is constructed,and then the constructed atmospheric dissipation function is modified according to the texture information to repair the transmission adaptivly.The experimental results show that the improved algorithm can remove the halo effect and solve the problem of bright region distortion.(2)Because of the low recognition accuracy of traffic signs by traditional recognition algorithms,a neural network method is introduced to recognize traffic signs.In order to solve the problem that VGG16 requires fixed input image size,VGG16 is improved by combining transfer learning knowledge and spatial pyramid pooling.In order to maintain the output of fixed dimension,the last layer of convolution feature map is pooled at multiple scales,and then the pooled results are merged.Experimental results show that the accuracy of improved VGG16 network can reach 97.8%.(3)The special TSD-CNN network is constructed by using Inception series of network ideas and combining the characteristics of traffic sign images.In order to solve the problem of different sizes of network image input,GAP is used instead of the full connection layer.Inception convolution module is introduced to extract more target features and increase network depth.The BN layer is introduced to accelerate the convergence speed of the network.At the same time,the Leaky ReLu activation function with better performance was used tosolve the problem that some neurons could not be activated.Experimental results show that the recognition accuracy of TSD-CNN is significantly improved,reaching more than 98%.