Research And Implementation Of Defects Statistic Method In Material Microscopic Image Based On Object Detection

In order to study the relationship between microstructure and macroscopic properties of materials,material experts need to statistically analyze the number and size of microstructure(such as defects,martensite,?-clusters)in electron microscope images.In this paper,the statistical problem of defects in material microscopic images is studied.The defects mainly include trapped bubbles and unfused compounds.At present,the defect statistics are done manually,generally,the electron microscope pictures of a group of materials will take 2 days,and it is very error-prone.The use of artificial intelligence instead of manual completion of this task can obtain accurate statistical data in a few minutes,which is of great significance to liberate experts and improve accuracy.Different from the dense cells with cross-movement,the defects in material micro-images are usually sparse and the number is small.In view of their sparseness,this paper proposes a new material microdefect statistical method based on object detection,instead of using the statistical method based on density map estimation,which is commonly used in micro-images.Defects in material micro-imgages are usually relatively small,large scale span,irregular shape,and the ratio of foreground to background is seriously unbalanced.These characteristics increase the difficulty of defect detection,and these problems must be solved or alleviated one by one.Firstly,based on Faster R-CNN,dilated convolution and featurized pyramid model are introduced into model design for small object problem and multi-scale problem,deformable convolution is introduced into model design for irregular shape,and focal loss function is introduced for the imbalance between foreground and background,so that the effect of the new model is improved from 0.460 to 0.539.At the same time,because defects are relatively few in material microimages,the over-fitting problem caused by a small amount of data has become another challenge to build and train a good and stable material micro-defect detection model.In this study,transfer learning and automatic data augmentation are used to solve the problem of overfitting and enhance the generalization ability of the model.Among them,the automatic data augmentation method will search out the data augmentation strategies suitable for material micro-images.On the basis of the improved model,through the improvement of the above training strategy,the average accuracy of the final model reaches 0.552.The experimental results show that the new method is effective.