Research On Defect Detection Algorithm Based On Convolutional Neural Network Compression

In recent years,the rapid growth of data volume and the breakthrough of computer hardware have enabled convolutional neural networks to be successfully applied in various industries.With the increasing requirements for detection and recognition accuracy in the field of industrial detection and identification,the convolutional neural network improves the feature extraction ability and the learning ability of the network by continuously increasing the depth of the network and the number of feature maps.Thereby achieving the requirements of detection accuracy.However,the network model only has high detection accuracy can not guarantee that the model can be deployed to the industrial production pipeline to detect and identify the workpiece in the actual application process.Inspecting defects on industrial production lines not only considers the accuracy of the network model,but also the real-time nature of inspections on industrial production lines and the computational overhead costs of hardware and software.The deep convolutional neural network model has a large size,which not only has a high memory requirement for the training model,but also has a large calculation amount for detecting the category of a single picture.Thus it limits the efficiency of detection and recognition.At the same time,considering that a large number of labeled sample data sets are needed in the network training process and the workpiece defects on the industrial production line are complicated.The labeling of the workpiece requires not only professional judgment,but also a large labor cost.Therefore,in this paper,the convolutional neural network model is compressed for the problem of excessive parameter quantity and computational complexity in the convolutional neural network model,and the improved active learning method is used to solve the problem ofdifficult marking of the workpiece data set in the training process.The network can be deployed to industrial production lines to detect defects without loss of precision.The main research contents of this paper are as follows:(1)In this paper,the basic structure of convolutional neural networks is studied first,and the advantages and disadvantages of convolutional neural network models are expounded.Aiming at the problem of excessive parameter quantity and computational complexity in deep convolutional neural network model,this paper proposes a targeted channel pruning method to sparse the deep convolutional neural network model.The network model can improve the recognition accuracy of the network model by performing fine adjustment after multiple pruning,and can compress the storage parameter amount of the network model and the calculation amount at the runtime.(2)Aiming at the problem that the parameter optimization in convolutional neural network pruning has high requirements on experience,a lightweight convolutional neural network model is proposed to detect and identify defective workpieces.The proposed lightweight convolutional neural network model reduces computational and parametric quantities by optimizing convolution calculations before model training,and no additional manual intervention is required after training is completed.It is verified by experiments that the lightweight convolutional neural network model can be directly deployed on the industrial production pipeline with high precision and low consumption detection performance.(3)Aiming at the problem that a large number of data sets need to be labeled during convolutional neural network training,this paper proposes an improved active learning algorithm.The algorithm combines two active selection strategies to select uniform samples in the training process of the network model.And then re-label the sample and add it to the labeled data set,iteratively train the network model until the network model achieves the required accuracy.The experiment verifies that the active learning algorithm proposed in this paper has a good performance in the convergence speed and labor cost of the model.