Research On Detection And Identification Of Paper Surface Defects Based On Convolutional Neural Network

With the rapid development of China's manufacturing industry,people's requirements for product quality are constantly improving,but defects are inevitable in the production process.Paper defects not only have an impact on the appearance itself and affect the use,so manufacturing enterprises pay special attention to product quality inspection,in order to ensure the quality of products,at the same time to analyze the quality inspection results,further improve the production process,reduce the generation of defects.At present,there are enterprises using machine vision technology paper defect detection system instead of manual quality inspection,but the algorithm can not meet the requirements of the paper machine,this paper paper surface defect detection and classification algorithm was studied,in order to find a suitable defect detection algorithm.The thesis mainly includes the following aspects:For the paper surface defect recognition,through the analysis of the paper surface in each link of the defect types and defect recognition and classification algorithms.Combined with the current convolution neural network,analyses the existing surface defect recognition algorithm research,by analyzing the mechanism of spatial attention attention information fusion channels,network is not limited to the original image feature extraction stages,put forward is added in the containing EffcientNet channel attention network space,design new attention mechanism model.The paper surface defects are classified.EffcientNet has a 0.65% higher accuracy than the original one,and its network structure and algorithm speed are not affected by adding the spatial attention mechanism.In terms of paper surface defect detection,the previous application of convolutional neural network in paper surface defect detection is analyzed and summarized,aiming at the problem that SSD(Single Shot Multibox Detector)has insufficient ability in detecting small targets.In this paper,an improved SSD algorithm is proposed.The efficient Net algorithm with attention module is used as the backbone network of SSD network to improve the target representation ability and enhance the detection ability.Due to the lack of data samples in product testing in this subject,the use of pre-training weight can accelerate gradient convergence,and achieve the training of small data sets,fine tuning of the network,and improve the problem of excessive fitting caused by small data and gradient disappearance and gradient explosion caused by improper initialization.Finally,due to the imbalance between positive samples and negative samples,the SSD network is difficult to conduct uniform and intensive sampling training for samples,which reduces the accuracy of the model.The focus loss function is used to mine the difficult samples and reduce the weight of simple negative samples in the training process to prevent the failure of detector in the training process.To further improve the accuracy of defect detection,the improved SSD network uses the feature pyramid to integrate the context information and rebuilds the pyramid feature mapping.Compared with the improved SSD network using EffcientNet,the Mean Average Precision of the model improves by 6%.The time was shortened by 5.2ms.