Research Based On Transfer Learning For Defect Detection Algorithms

In order to solve the problem of insufficient sample data for industrial surface defect detection,we propose an improved domain-adversarial deep learning neural network for defect detection of magnetic tile data,and the results are better than those of the existing mainstream network.The main contents of this paper are as follows:Firstly,in the field of industrial automation,the defect detection,as an important research direction,has always been the key of research in many industrial vision companies.At present,the main method of the defect detection is to use traditional detection algorithm to detect products by a camera.Traditional methods of machine vision use a machine learning classifier by extracting manual features,which is of low accuracy,is not robust to environment and needs manual correction.In recent years,deep learning has been developing rapidly in the field of computer vision.With the advantages of high accuracy and strong robustness,deep learning has become the mainstream of surface defect detection in industrial.However,deep learning technology for surface defect requires a large number of training samples with highquality labeling.In reality,obtaining high quality data with labels cost too much,which has become the bottleneck for the development of deep learning technology in the field of industrial surface defect detection.Secondly,in order to solve the problem of deep learning,we study the existing deep migration network in the paper.It includes finetune method,adaptive layer and adversarial learning.We make a detailed analysis and comparison of these networks,tests and compare the task data of defect detection,and point out the existing network problems in this paper.Nextly,in order to solve the problem,we propose a detection algorithm for defect detection based on transfer learning in this paper.We design an end-to-end improved domain-adversarial training of neural networks to to solve the problem of surface defect detection with insufficient training samples.Compared with domainadversarial training of neural networks,we use improved methods including category-based domain adaptation,weighted learning of samples and A-softmax loss function.In order to get a better model,we also explores the structure of feature extractor in this paper.Lastly,in order to evaluate the above models,we use DAMG 2007 data set as the source domain,magnetic tile data as the target domain for training,and compare our work with the existing mainstream network for transfer learning.Experiments show that under the same super-parameters,our model achieve the best performance with the advantage of 10 percentage points higher than other models in the test set.In order to observe the performance of each module for our model,we train and evaluate the networks which drops one improvement module separately.The experiment shows that each improvement module can promote the improvement of the model.The sample transfer method,category-based domain adaptation method and the use of A-softmax loss function in this paper are very helpful to improve the performance of the model.At the same time,these improved techniques can also be extended to other transfer tasks.