An improved surface defect classification algorithm for hot-rolled steel strips was presented based on the lightweight neural network MobilenetV3-large.In order to quickly reduce the number of parameters,pruning was used,greatly reducing the number of convolution layers,adjusting the channel size and step size,and modifying the corresponding network parameters.In order to compensate for the decline in accuracy caused by the modification of the model,the activation function ReLU was modified to the Hard-Swish,and the shuffle attention mechanism was introduced to replace the squeeze-and-excita-tion attention mechanism in the original model to further reduce the number of parameters while improv-ing the operating efficiency and classification accuracy.The experimental results showed that the parame-ters of the improved algorithm were 0.5 MB,96.89％less than the original model,the time spent training a picture reduced from 19.81 ms to 10.73 ms,and the average accuracy of the NEU-CLS surface defect dataset was 99.26％,being 5.56％higher than before the improvement and indicating that the improved algorithm can be applied to real-time classification.
维普
万方数据
