| Printing is not only a means of production and a tool for cultural communication,but also one of the basic industries essential to modern industrial production.The printing industry covers a wide range of fields,such as image processing,design,printing material production,and equipment manufacturing.Rollers are one of the important equipment in printing production,and their surface finish directly affects the quality of printed products,which has a great impact on the production efficiency and reputation of the whole printing industry.The inspection about the surface quality of printing rollers is getting more and more attention from the industry.With the progress of science and technology,how to quickly and effectively identify the defects on the surface of the roller has become a research hotspot,and for this problem,the research of this thesis includes the following points:(1)For the problem that the defect images on the surface of printing rollers exist in different sizes and have a large number of defects,a salient object detection method based on stepwise multi-scale feature fusion is proposed,which consists of a stepwise multi-scale feature guidance module,a multi-resolution information fusion module and a deeply supervised side output.Specifically,the stepwise multiscale feature guidance module contains two paths,top-down and bottom-up,to enhance the transfer of multiscale features in the network,and the multi-level features are fused by the multi-resolution information fusion module to suppress the interference of background noise and redundant features.In addition,deeply supervised side outputs are used to eliminate the discrepancy caused by background and foreground imbalance.The experimental results on the roller surface defect dataset show that this study can achieve0.7959,0.0144 in the two evaluation indexes of Fmax,MAE respectively,and the detection rate is 24.98 fps,which has better detection performance compared with other algorithms.(2)A visual saliency-based classification method for printing roller surface defects is proposed to address the problems of low accuracy and efficiency of printing roller surface defects classification.The proposed method suppresses the high-frequency components in the background texture by traditional visual saliency algorithm and deep information fusion algorithm,fully extracts the image feature information by using multi-group convolutional parallel structure to enhance the multi-scale expression capability of the network and improve the classification performance,and retains the negative information in the inference process by the activation function PRe LU to improve the nonlinear expression capability of the network.The experimental results show that the method can effectively distinguish the surface defects of printing rollers with the classification accuracy of 98.50% and the average recognition speed of 64.10 fps,which basically meets the production requirements of the printing industry. |
