As the demand for materials with excellent mechanical properties is increasing in scientific research and engineering, determining how to accurately measure the global displacement field of materials in mechanical experiments has become an important scientific research issue. Digital image correlation (DIC) algorithm is a non-contact optical method for measuring global speckle displacement fields based on visible light, which is widely used in experimental mechanics and engineering fields. It has the advantages of low measurement costs, high precision, high sensitivity, strong anti-interference ability, and global measurement. However, traditional DIC algorithm cannot meet the requirements of real-time measurement in practical applications, which greatly limits the development and promotion of this method. With the rapid development of deep learning in computer vision, deep learning methods gradually come into use in DIC algorithm. Thanks to the efficient calculation by general processing unit (GPU) devices, the deep learning-based method for measuring the speckle displacement field can more easily achieve real-time online calculation. Although the method is much faster than the traditional one, the model cannot accurately measure the complex large deformation displacement field in practical applications due to the incomplete dataset. Hence, this work aims to construct a more realistic and comprehensive speckle image dataset with a large deformation displacement field and propose a fast and high-precision deep learning model to measure the displacement field of speckle images with large deformation.
A large number of different types of speckle images is obtained in various ways (Fig. 1) to construct a speckle image dataset with a large deformation displacement field in line with the actual situation. These speckle images are obtained from real experiments and computer simulations under different parameter combinations (Table 1). Then, a composite deformation composed of translation, stretching, compression, rotation, Gauss, shear, and other basic deformations is used to define the random displacement field. Finally, a speckle-image displacement-field dataset with a maximum displacement of 16 pixel and large deformation in line with the actual deformation is produced. In terms of the deep learning network model, a fast and high-precision network model DICNet (Fig. 5) for measuring the speckle images with a large-deformation displacement field is built upon the improvement on UNet. DICNet introduces a convolutional block attention module to increase the efficiency of feature extraction and fusion, uses depthwise separable convolution to replace some ordinary convolutional layers, and increases the convolution kernel size of some convolutional layers. It improves the displacement-field measurement accuracy and reduces the number of parameters of the network model. At the network training stage, a combination of the global shape loss function and global absolute loss function is proposed to improve the convergence speed and accuracy of the model.
Network selection experiments are conducted to prove that UNet is a rational basic network model for measuring the large-deformation displacement (Table 2). It has higher measurement accuracy of the displacement field, a smaller number of parameters, and faster inference speed. The DICNet proposed in this work is compared with the traditional DIC algorithm and the latest deep learning methods on the self-built dataset, and the performance of these methods in the measurement task of the large-displacement displacement field is comprehensively compared in terms of three indicators, i.e., the root-mean-square error (RMSE), the standard deviation, and mean time (Table 3). The results show that the measurement accuracy of the deep learning method is better than that of the traditional method. The RMSE of DICNet on the training set and the validation set is 0.056 pixel and 0.055 pixel, respectively, which is 67%-70% lower than that of other existing methods and about 39% lower than that of the original UNet network. On the test set, DICNet still has the smallest RMSE and the most stable performance (Table 4). The experiments of DICNet are also conducted on the public DIC challenge dataset (Fig. 8). The results show that the measurement results of the proposed method are highly consistent with those of the traditional algorithms, which indicates that the proposed method still has good generalization performance on the public dataset.
This work proposes a displacement field measurement method for speckle images with complex large deformation. This method uses the convolutional block attention module and depthwise separable convolution to improve the UNet network for measuring large deformation displacement fields. To train the model, this work constructs a dataset containing multiple types of speckle images and complex large-deformation displacement fields in line with the real situation and proposes a new loss function. This method is compared with traditional DIC algorithm and the latest deep learning methods on the self-built dataset and public dataset separately. The results show that the measurement results of DICNet are highly consistent with those of other methods, and the method in this work achieves the highest average accuracy with the smallest number of model parameters. The measurement speed of the displacement field is far higher than those of traditional methods, which can meet the actual real-time measurement requirements of a large deformation displacement field. The source code and network pre-trained weights of this study are available at