| Accidents caused by material fracture failure emerge one after another,causing many serious losses.Studying the internal failure process of materials and revealing the internal damage evolution mechanism is crucial for preventing fracture failure.Computed tomography(CT)can provide a three-dimensional internal characterization of damages and provide an opportunity to study the internal damage evolution mechanism of materials.However,three-dimensional CT image data is large,and the damage evolution features are often small in structure and low in contrast,thus facing the challenge of quantitative recognition and extraction of damage evolution.Herefore,aim at the different damage features in the two key evolution stages of internal damage initiation and expansion in in-situ CT images,this paper proposes dedicated damage identification and extraction methods and verifies their effectiveness through experimental data.Specific research contents are as follows:(1)Identification of crack initiation stage.According to the crack features of small size and low contrast in the initiation stage,prior terms are introduced to enhance the features of tiny cracks while guiding and constraining the network training.In this paper,two applicable methods are proposed:Mask R-CNN method based on structural evolution and U-Net method based on digital volume correlation(DVC)strain priori guidance.The Mask R-CNN method based on structure evolution introduces the difference images which indicating structure evolution to correct network prediction results.Due to the corresponding relationship between strain and crack location,the strain field is introduced as input information in the U-Net method based on DVC strain priori guidance,which guides the network to consider the strain parameter of the crack.The two methods were tested on the helical laminated composites.From the perspective of image segmentation and mechanical parameter,it is verified that the segmentation performance of the two methods is improved compared with the corresponding basic model.Both methods can improve the precision rate of tiny cracks identification and reduce the error identification of non-crack regions.Among them,U-Net method based on DVC strain prior guidance has the best segmentation effect.(2)Tracking of crack expansion stage.According to the features of different crack shape and directions but high contrast in the expansion stage,two methods are proposed:Mask R-CNN based on crack shape constraint and U-Net network method combined with active contour loss function based on crack gray uniformity.The Mask R-CNN method based on crack shape constraints takes into account the feature that the crack shape is long and narrow.According to the statistical data of crack length to width ratio,the initial anchor frame ratio is replaced,which improves the training efficiency and increases the confidence of the target frame.The active contour loss function based on crack gray uniformity considers the region information,that is,the difference between the inner and outer regions is the largest and the gray difference between the inner regions of the crack is the smallest,which effectively reduces the influence of noise and can maintain the crack shape well.These two methods are tested on orthogonal laminated composite materials.From the perspective of image segmentation and mechanical parameter,it is verified that the segmentation performance of the two methods is improved compared with the corresponding basic model.Among them,the active contour loss function based on crack gray uniformity combined with U-Net network has the best segmentation effect.(3)Application and effect evaluation of identification and extraction methods in the whole process of damage evolution.On the basis of the artificial composite materials mentioned above,the internal damage identification method proposed in this paper is futher adopted for natural biological materials with complex structures in stages to verify the effectiveness of the method.At the crack initiation stage,the two methods proposed in this paper are used to qualitatively and quantitatively compare the segmentation effects of the two methods,which again proves the superiority of the UNet method based on DVC strain priori guidance in the identification of micro cracks.At the crack expansion stage,the two methods proposed in this paper are used to qualitatively and quantitatively compare the segmentation effects of the two methods,which again proves the superiority of the active contour loss function based on crack gray uniformity combined with U-Net network in maintaining crack shape.At each stage,the corresponding optimal recognition method is used to identify and extract damage,and the 2D segmentation results are stacked into 3D,which can more clearly show the whole process of damage from initiation to expansion. |
