Research On Improved Faster R-CNN Based Paper Defect Diagnosis Algorithms

During the process of papermaking,various factors can affect the surface of the paper,resulting in varying degrees of paper defects,such as wrinkles,black spots,holes,etc.The emergence of paper defects has a significant impact on the special-purpose paper used as aerospace paper,industrial capacitor paper,highend wallpaper base paper and coated paper,even causing safety hazards.Therefore,paper defect diagnosis is an indispensable link in the papermaking process for special papers.The process of paper defect diagnosis is divided into three links:paper image data collection,paper defect image information extraction and paper defect type recognition.Due to the continuous increase of paper width and paper machine speed,the number of collected paper images has also increased.At the same time,the probability of paper defects occurring has greatly increased,making it increasingly difficult to detect paper defects online.This thesis is focused on the improvement of paper defect diagnosis algorithms in paper defect diagnosis process,which address issues such as insufficient feature extraction,low detection accuracy,missed detections and poor real-time performance in paper defect images.The main work of the thesis can be summarized into the following four aspects.Ⅰ.Research on paper defect detection algorithm based on improved Faster RCNNA paper defect detection algorithm based on improved Faster R-CNN is proposed to address the issues of insufficient feature extraction,low detection accuracy and missed detection of small targets in traditional paper defect detection algorithms.The improvement lies in the dual attention mechanism CBAM added space and channel in the algorithm which improves the accuracy of paper defect detection.Furthermore,ROI-Pooling is replaced by ROI-Align to enhance the generalization ability of the network.The experimental results show that the average accuracy of the improved algorithm reaches 98%,which is 8.5%higher than the original algorithm.The paper defect detection algorithm based on improved Faster R-CNN can fully extract the feature information of paper defects,effectively improve the detection accuracy of small target paper defects,and reduce the error and missed detection rates.Ⅱ.Research on model compression based on improved Faster R-CNNAiming at the problems of complex Network structure and slow detection speed of ResNet-50 in the improved Faster R-CNN model,an improved Network Sliming model compression algorithm is proposed.The improvements are as follows:ELU activation function is used instead of ReLU activation function to alleviate the death of neural nodes in the model;BN layer and ELU activation function are placed before the convolution operation to improve the generalization ability of the model.Experimental results show that the detection speed of the improved Faster R-CNN model is increased by 20.8%compared with that before compression.Ⅲ.Construction of an experimental platform for paper defect diagnosis and validation of algorithm examplesTo verify the feasibility of paper defect detection algorithms,a corresponding experimental platform is established to provide experimental data and a testing environment for algorithm research.The experimental platform includes a light source module,a paper image acquisition device,a paper diagnosis module and other parts.Manufacturing paper defects on the experimental platform,selecting a suitable light source to illuminate the paper,using a CCD camera to capture paper images and transmitting the image information to the paper defect diagnosis module for diagnosis.The paper defect diagnosis algorithm proposed by the improved Faster R-CNN after compression is verified.The experimental results show that the average detection accuracy of the improved algorithm can reach 98.7%,effectively improving the detection accuracy of paper defect.The detection speed of the improved algorithm is increased by 20.6%,and the real-time performance of paper defect diagnosis is improved.Ⅳ.Visualization implementation of paper defect diagnosis systemIn order to realize the visualization of paper defect detection results,a paper defect diagnosis system based on PyQt5 development tool is designed to visualize the paper defect diagnosis process.The system designs a user interface library based on the actual requirements of the paper defect diagnosis system and implements functions within user management,paper defect detection,and paper defect recognition and classification.For testing the accuracy,100 images from the four types of paper defects are randomly selected for functional testing in the system.This testing platform provides a new idea for the practical research of subsequent industrial products.In summary,the main work of this paper is based on the improved Faster RCNN paper disease diagnosis algorithm.The experimental results show that the average accuracy of the improved algorithm can reach 98%and the detection speed increases by 20.8%.Moreover,a corresponding experimental platform is established to validate the algorithm by establishing a paper defect dataset.Further more,a paper defect diagnosis system software is designed to achieve visualization of paper defect detection and recognition classification results,laying the foundation for subsequent online paper defect diagnosis.