Abstract: The defect characteristics of metal cylindrical workpieces are easily affected by light.However,the traditional convolutional neural network which is used to detect the defects of metal cylindrical workpieces has many disadvantages and limitations,such as large parameters,great computational complexity,and a bad generalization.Besides,it cannot meet the requirements of high real-time performance and accuracy in the application of industrial detection.Thus,a lightweight network which incorporates the example-batch normalization network(IBN-NET)was proposed.Based on the lightweight convolutional neural network SqueezeNext,a basic module of the proposed network was constructed by adding IBN-NET which has a good generalization,and replacing the batch normalization(BN)after the shallow convolutional layer with a certain proportion of instance standardization(IN).Then,an improved network model was built by assembling the above basic modules.In the experiments,five types of defects of the metal cylindrical workpieces were used for evaluating the performance of the proposed model.The experiment results showed that the proposed model which incorporates IBN-NET had better generalization.Specifically,the proposed model has only 0.58 million parameters and could finish detecting one metal cylindrical workpieces using 5.54 ms with the precision of 95.8%when running in the graphics card of GTX1080.