Image Defect Recognition Based On Convolutional Neural Network For Ultrasonic Nondestructive Testing

With the development of the country's industrialization,the manufacturing industry is the main force of the industry,and the quality and safety of material products have also caused people's attention.There are various kinds of material products in life.When there are defects in material products,it will have a great impact on people's lives,not only cause waste,but also seriously cause life safety issues.Therefore,quality inspection of material products is required.Ultrasonic testing,as a convenient,safe and harmless testing method,is often used for defect detection.Most of the ultrasonic testing results obtained are only manually identified by experience,which is not conducive to industrial intelligent development and affects efficiency.In order to solve this problem,based on the knowledge theory of machine learning and deep learning,this paper conducts research on the identification of ultrasonic image defects.The purpose is to realize the automatic identification of ultrasonic image defects.The research content mainly includes the following parts:(1)Image feature processing based on the combination of PCA and Relief F algorithm to complete image defect recognition.Ultrasonic image defect recognition,manual recognition is low in efficiency and prone to misjudgment.Based on the characteristics of the collected ultrasound images,this paper analyzes the diversity and importance of features under different targets,combines image texture,shape,and spatial information to perform feature extraction,uses PCA-Relief F algorithm to combine feature reduction and selection,and uses BP neural network for images Identify.The experimental data shows that after combining the PCA and Relief F algorithms,the recognition accuracy rate reaches78.3%,which improves the recognition efficiency and accuracy rate.(2)Aiming at machine learning methods,it requires human participation,judging image features,selecting and extracting appropriate features,which can easily affect the recognition results,and requires tedious steps such as step-by-step operations.A small convolution network is proposed for end-to-end ultrasound image defect recognition.The model uses raw ultrasound images for image data enhancement,and uses convolution for feature extraction.Combining feature extraction and recognition to achieve end-to-end recognition,the recognition results of different depth models are compared.The experimental comparison shows that the small convolutional network can complete the image defect recognition,simplify the image recognition steps,and perform better in terms of recognition accuracy.(3)Aiming at the small number of image data samples and prone to overfitting of convolutional networks,a method combining deep convolutional networks and transfer learning is proposed.This method uses a deep network model trained on a large-scale match dataset to construct a fully connected classification layer of the deep convolutional network model.The convolutional feature extraction layer freezes the trainable parameters and the model performs training verification.The deep convolutional network feature extraction is performed.Fine-tune the parameters and experiment.Experimental verification shows that the model recognition accuracy after migration fine-tuning reaches98.89%.Through the freeze training of different models with different convolutional layers,Inception V3 and Res Net50 have higher recognition accuracy when the model parameter ratio is about 75%.The experimental results also show that after the method of fine tuning of migration,a deep convolution model with better recognition performance for small sample target tasks can be obtained.Finally,based on the self-built ultrasonic image defect data set,the model method recognition application is summarized and selected to obtain the optimal model for ultrasonic image defect recognition,and the Inception V3 model under migration fine-tuning is obtained.For the convolutional network model,the gradient inverse calculation is used to realize the visualization of the features extracted by the convolutional layer,which intuitively shows the feature extraction and transfer of features by different convolutional layers of the network model.