Research On Anomaly Detection Algorithm For HRTEM Crystal Image

Crystal is an important material in the field of materials science.Its mechanical properties are affected by the alignment of internal particle.If the alignment of the internal particles is deviated from the regular position,it is called a crystal defect.The high-resolution transmission electron microscope(HRTEM)image is one of the important data in the study of crystal defects.In the defect analysis tasks,to facilitate the subsequent processing,an abnormal detection stage is always carried out to the HRETM images firstly,to distinguish the images of abnormal crystals from those the normal ones.Deep learning related theories and technologies have made great achievements in the traditional field of image processing,but its application in anomaly detection for HRTEM crystal images is relatively few.To improve anomaly detection performance,deep learning based HRTEM crystal image anomaly detection is researched in this paper,and the main contributions are as follows:A novel HRTEM crystal image anomaly recognition algorithm based on Selective Kernel Networks(SKNet)is proposed.The variable size convolution kernel of SKNet is utilized to improve the scale robustness of the algorithm.For the problem of the loss of corresponding relation between channel and weight caused by data compression in SKNet,an Efficient Channel Attention Module(ECA Module)is introduced to replace the Squeeze and Excitation Block(SE Block),and the interaction of information across channels is obtained without reducing dimension.In addition,the focal loss is used to replace the cross entropy loss in the original network,which enhances the classification performance of the algorithm for difficult samples.Experimental results show that compared with the original SKNet network,the proposed algorithm improves the average classification accuracy of normal and abnormal samples by 3.9%,the classification precision for abnormal samples by 1.5%,and the recall rate by 5.2%.Furthermore,for the problem of imbalance between normal crystal image and abnormal crystal image samples,the semi-supervised abnormal detection method is studied,and a HRTEM image anomaly detection algorithm based on Patch Support Vector Data Description(Patch SVDD)is proposed.Firstly,the Cut Paste method is used to construct pseudo-abnormal samples,and have contrastive learning with normal samples to improve the discriminatory ability of the algorithm between normal samples and abnormal ones.Secondly,the feature extraction module of Patch SVDD is improved.The extracted shallow features are fused with deep features to improve the information carried in the feature and improve the detection performance.Finally,for the detection error caused by the crystal boundary region,the crystal image is segmented using Deep Lab V3+,and the influence of the boundary region on the detection result is eliminated by setting the abnormal score of the non-crystal region to zero.The experimental results show that compared with the original Patch SVDD,the AUROC value of the proposed algorithm is improved by 8.1%.