基于改进ShuffleNet V2网络的路面类型识别OA

To solve the problems of pavement type recognition models with large volume and low accuracy,the improved ShuffleNet V2 network pavement type recognition model was proposed.The efficient channel attention(EC A)module was added to the ShuffleNet V2 network structure to achieve cross channel information interaction by attention mechanism.The size of convolution kernel was adjusted according to the number of input channels.The ReLU function was replaced by the LeakyRelu function to avoid the invalidation of activation function.In order to improve the feature extraction ability and generalization ability of the model,the module composed of inflated convolution was introduced to obtain the wider range of image information with the image resolution unchanged.According to the classification characteristics of pavement types,the number of each module stacked and the overall architecture of the network were adjusted to reduce the model's computational and parametric quantities.The improved algorithm was verified on the road surface classification dataset(RSCD).The results show that the enhanced ShuffleNet V2 model achieves parameter quantity of 4.67×106,representing reduction of 1.4 ×105 compared to the original model.The accuracy reaches 95.53％with improvement of 0.71 percentage point over the pre-optimization level.The inference time is reduced by 31％,and the accuracy of road surface type recognition and response speed are improved.