Research On Infrared Defect Detction Based On Deep Learning Spatio-temporal Feature Extraction

Carbon Fiber Reinforced Polymer/Plastic(CFRP)has some good properties such as low density,however,it has a possibility to produce defects such as debonding and fracture during long-term service,which brings potential safety hazards to equipments.Infrared non-destructive testing technology on the basis of light excitation has benefits of non-contact,safety,fast speed,and easy operation,which is often applied to the detection of internal defects in composite materials.However,the collected thermal videoes,based on the nature of the materials and equipments themselves,are prone to be interfered by high background noise,etc.,and defects cannot be observed directly,therefore,it is necessary to develop defect detection algorithms.Deep learning algorithms can perform feature extraction on video sequence data,and can automatically capture and recognize target information through self-learning.In infrared detection field,due to the limitations of infrared thermal imaging data such as background noise,light interference,weak targets,and few samples,current detection algorithms have challenges such as poor generalization ability and inability to accurately and comprehensively locate defects.In this master thesis,through the study of deep learning algorithm models and algorithms in the field of image processing,in-depth mining and analysis of light-excited infrared thermal imaging video data,fusion of multi-dimensional convolutional neural networks and spatio-temporal structure algorithm,we have proposed a defect detection automatic recognition model.Innovated from data processing,model and inspection efficiency,etc.,we propose a new end-to-end automatic defect detection algorithm.The main work of this thesis is as follows:1)Two different light-excited infrared thermal imaging systems,fixed and portable,are used to test multiple types of specimens repeatedly,obtain video data of infrared thermal imaging defects,and establish an infrared imaging defect detection database.According to the characteristics of thermal imaging defect data,it analyzes and processes,reduces noise in the preprocessing stage,improves the network model,and establishes a video image defect semantic segmentation and detection system based on spatio-temporal feature extraction.2)This thesis has proposed an end-to-end infrared non-destructive inspection model grounded on feature extraction and fusion on different dimensions.This method is based on the deep learning image segmentation model and traditional feature extraction methods,and extracts and merges defect features from spatio-temporal perspective.The self-attention mechanism is used to make the model adaptively learn the typical characteristics of defects.Moreover,the network model parameters is decreased to lower the detection time,finally,a lightweight network infrared thermal defect detection algorithm is achieved.In different test systems,different shapes and types of test specimens,the effectiveness and robustness of the algorithm are verified and tested,the defect detection effects of various semantic segmentation algorithms are objectively compared.The results show that the defect detection algorithm proposed in this paper has better defect detection performance.